Started unit tests for the review controller (#15077)

* Started unit tests for the review controller

* Revert "Started unit tests for the review controller"

This reverts commit 7746eb146f.

* Started unit tests for the review controller

* FIrst test

* Added test for review endpoint (time filter - after + before)

* Assert expected event

* Added more tests for review endpoint

* Added test for review endpoint with all filters

* Added test for review endpoint with limit

* Comment

* Renamed tests to increase readability

.cspell/frigate-dictionary.txt

@@ -12,6 +12,7 @@ argmax
 argmin
 argpartition
 ascontiguousarray
+astype
 authelia
 authentik
 autodetected
@@ -42,6 +43,7 @@ codeproject
 colormap
 colorspace
 comms
+coro
 ctypeslib
 CUDA
 Cuvid
@@ -59,6 +61,7 @@ dsize
 dtype
 ECONNRESET
 edgetpu
+fastapi
 faststart
 fflags
 ffprobe
@@ -193,6 +196,7 @@ poweroff
 preexec
 probesize
 protobuf
+pstate
 psutil
 pubkey
 putenv
@@ -212,6 +216,7 @@ rcond
 RDONLY
 rebranded
 referer
+reindex
 Reolink
 restream
 restreamed
@@ -236,6 +241,7 @@ sleeptime
 SNDMORE
 socs
 sqliteq
+sqlitevecq
 ssdlite
 statm
 stimeout
@@ -270,9 +276,11 @@ unraid
 unreviewed
 userdata
 usermod
+uvicorn
 vaapi
 vainfo
 variations
+vbios
 vconcat
 vitb
 vstream

.devcontainer/post_create.sh

@@ -3,10 +3,12 @@
 set -euxo pipefail
 
 # Cleanup the old github host key
-sed -i -e '/AAAAB3NzaC1yc2EAAAABIwAAAQEAq2A7hRGmdnm9tUDbO9IDSwBK6TbQa+PXYPCPy6rbTrTtw7PHkccKrpp0yVhp5HdEIcKr6pLlVDBfOLX9QUsyCOV0wzfjIJNlGEYsdlLJizHhbn2mUjvSAHQqZETYP81eFzLQNnPHt4EVVUh7VfDESU84KezmD5QlWpXLmvU31\/yMf+Se8xhHTvKSCZIFImWwoG6mbUoWf9nzpIoaSjB+weqqUUmpaaasXVal72J+UX2B+2RPW3RcT0eOzQgqlJL3RKrTJvdsjE3JEAvGq3lGHSZXy28G3skua2SmVi\/w4yCE6gbODqnTWlg7+wC604ydGXA8VJiS5ap43JXiUFFAaQ==/d' ~/.ssh/known_hosts
-# Add new github host key
-curl -L https://api.github.com/meta | jq -r '.ssh_keys | .[]' | \
-  sed -e 's/^/github.com /' >> ~/.ssh/known_hosts
+if [[ -f ~/.ssh/known_hosts ]]; then
+  # Add new github host key
+  sed -i -e '/AAAAB3NzaC1yc2EAAAABIwAAAQEAq2A7hRGmdnm9tUDbO9IDSwBK6TbQa+PXYPCPy6rbTrTtw7PHkccKrpp0yVhp5HdEIcKr6pLlVDBfOLX9QUsyCOV0wzfjIJNlGEYsdlLJizHhbn2mUjvSAHQqZETYP81eFzLQNnPHt4EVVUh7VfDESU84KezmD5QlWpXLmvU31\/yMf+Se8xhHTvKSCZIFImWwoG6mbUoWf9nzpIoaSjB+weqqUUmpaaasXVal72J+UX2B+2RPW3RcT0eOzQgqlJL3RKrTJvdsjE3JEAvGq3lGHSZXy28G3skua2SmVi\/w4yCE6gbODqnTWlg7+wC604ydGXA8VJiS5ap43JXiUFFAaQ==/d' ~/.ssh/known_hosts
+  curl -L https://api.github.com/meta | jq -r '.ssh_keys | .[]' | \
+    sed -e 's/^/github.com /' >> ~/.ssh/known_hosts
+fi
 
 # Frigate normal container runs as root, so it have permission to create
 # the folders. But the devcontainer runs as the host user, so we need to

.github/DISCUSSION_TEMPLATE/detector-support.yml

@@ -74,19 +74,6 @@ body:
         - CPU (no coral)
     validations:
       required: true
-  - type: dropdown
-    id: object-detector
-    attributes:
-      label: Object Detector
-      options:
-        - Coral
-        - OpenVino
-        - TensorRT
-        - RKNN
-        - Other
-        - CPU (no coral)
-    validations:
-      required: true
   - type: textarea
     id: screenshots
     attributes:

.github/DISCUSSION_TEMPLATE/general-support.yml

@@ -102,19 +102,6 @@ body:
         - CPU (no coral)
     validations:
       required: true
-  - type: dropdown
-    id: object-detector
-    attributes:
-      label: Object Detector
-      options:
-        - Coral
-        - OpenVino
-        - TensorRT
-        - RKNN
-        - Other
-        - CPU (no coral)
-    validations:
-      required: true
   - type: dropdown
     id: network
     attributes:

.github/pull_request_template.md

@@ -13,6 +13,7 @@
 - [ ] New feature
 - [ ] Breaking change (fix/feature causing existing functionality to break)
 - [ ] Code quality improvements to existing code
+- [ ] Documentation Update
 
 ## Additional information
 

.github/workflows/ci.yml

@@ -6,6 +6,8 @@ on:
     branches:
       - dev
       - master
+    paths-ignore:
+      - 'docs/**'      
 
 # only run the latest commit to avoid cache overwrites
 concurrency:
@@ -155,6 +157,28 @@ jobs:
             tensorrt.tags=${{ steps.setup.outputs.image-name }}-tensorrt
             *.cache-from=type=registry,ref=${{ steps.setup.outputs.cache-name }}-amd64
             *.cache-to=type=registry,ref=${{ steps.setup.outputs.cache-name }}-amd64,mode=max
+  arm64_extra_builds:
+    runs-on: ubuntu-latest
+    name: ARM Extra Build
+    needs:
+      - arm64_build
+    steps:
+      - name: Check out code
+        uses: actions/checkout@v4
+      - name: Set up QEMU and Buildx
+        id: setup
+        uses: ./.github/actions/setup
+        with:
+          GITHUB_TOKEN: ${{ secrets.GITHUB_TOKEN }}
+      - name: Build and push Rockchip build
+        uses: docker/bake-action@v3
+        with:
+          push: true
+          targets: rk
+          files: docker/rockchip/rk.hcl
+          set: |
+            rk.tags=${{ steps.setup.outputs.image-name }}-rk
+            *.cache-from=type=gha
   combined_extra_builds:
     runs-on: ubuntu-latest
     name: Combined Extra Builds

.github/workflows/pull_request.yml

@@ -1,6 +1,9 @@
 name: On pull request
 
-on: pull_request
+on:
+  pull_request:
+    paths-ignore:
+      - 'docs/**'
 
 env:
   DEFAULT_PYTHON: 3.9

.github/workflows/release.yml

@@ -34,14 +34,14 @@ jobs:
           STABLE_TAG=${BASE}:stable
           PULL_TAG=${BASE}:${BUILD_TAG}
           docker run --rm -v $HOME/.docker/config.json:/config.json quay.io/skopeo/stable:latest copy --authfile /config.json --multi-arch all docker://${PULL_TAG} docker://${VERSION_TAG}
-          for variant in standard-arm64 tensorrt tensorrt-jp4 tensorrt-jp5 rk; do
+          for variant in standard-arm64 tensorrt tensorrt-jp4 tensorrt-jp5 rk h8l rocm; do
             docker run --rm -v $HOME/.docker/config.json:/config.json quay.io/skopeo/stable:latest copy --authfile /config.json --multi-arch all docker://${PULL_TAG}-${variant} docker://${VERSION_TAG}-${variant}
           done
 
           # stable tag
           if [[ "${BUILD_TYPE}" == "stable" ]]; then
             docker run --rm -v $HOME/.docker/config.json:/config.json quay.io/skopeo/stable:latest copy --authfile /config.json --multi-arch all docker://${PULL_TAG} docker://${STABLE_TAG}
-            for variant in standard-arm64 tensorrt tensorrt-jp4 tensorrt-jp5 rk; do
+            for variant in standard-arm64 tensorrt tensorrt-jp4 tensorrt-jp5 rk h8l rocm; do
               docker run --rm -v $HOME/.docker/config.json:/config.json quay.io/skopeo/stable:latest copy --authfile /config.json --multi-arch all docker://${PULL_TAG}-${variant} docker://${STABLE_TAG}-${variant}
             done
           fi

docker-compose.yml

@@ -23,7 +23,7 @@ services:
     #                     count: 1
     #                     capabilities: [gpu]
     environment:
-      YOLO_MODELS: yolov7-320
+      YOLO_MODELS: ""
     devices:
       - /dev/bus/usb:/dev/bus/usb
       # - /dev/dri:/dev/dri # for intel hwaccel, needs to be updated for your hardware

docker/hailo8l/Dockerfile

@@ -16,89 +16,25 @@ RUN mkdir /h8l-wheels
 # Build the wheels
 RUN pip3 wheel --wheel-dir=/h8l-wheels -c /requirements-wheels.txt -r /requirements-wheels-h8l.txt
 
-# Build HailoRT and create wheel
-FROM wheels AS build-hailort
+FROM wget AS hailort
 ARG TARGETARCH
-
-SHELL ["/bin/bash", "-c"]
-
-# Install necessary APT packages
-RUN apt-get -qq update \
-    && apt-get -qq install -y \
-    apt-transport-https \
-    gnupg \
-    wget \
-    # the key fingerprint can be obtained from https://ftp-master.debian.org/keys.html
-    && wget -qO- "https://keyserver.ubuntu.com/pks/lookup?op=get&search=0xA4285295FC7B1A81600062A9605C66F00D6C9793" | \
-    gpg --dearmor > /usr/share/keyrings/debian-archive-bullseye-stable.gpg \
-    && echo "deb [signed-by=/usr/share/keyrings/debian-archive-bullseye-stable.gpg] http://deb.debian.org/debian bullseye main contrib non-free" | \
-    tee /etc/apt/sources.list.d/debian-bullseye-nonfree.list \
-    && apt-get -qq update \
-    && apt-get -qq install -y \
-    python3.9 \
-    python3.9-dev \
-    build-essential cmake git \
-    && rm -rf /var/lib/apt/lists/*
-
-# Extract Python version and set environment variables
-RUN PYTHON_VERSION=$(python3 --version 2>&1 | awk '{print $2}' | cut -d. -f1,2) && \
-    PYTHON_VERSION_NO_DOT=$(echo $PYTHON_VERSION | sed 's/\.//') && \
-    echo "PYTHON_VERSION=$PYTHON_VERSION" > /etc/environment && \
-    echo "PYTHON_VERSION_NO_DOT=$PYTHON_VERSION_NO_DOT" >> /etc/environment
-
-# Clone and build HailoRT
-RUN . /etc/environment && \
-    git clone https://github.com/hailo-ai/hailort.git /opt/hailort && \
-    cd /opt/hailort && \
-    git checkout v4.18.0 && \
-    cmake -H. -Bbuild -DCMAKE_BUILD_TYPE=Release -DHAILO_BUILD_PYBIND=1 -DPYBIND11_PYTHON_VERSION=${PYTHON_VERSION} && \
-    cmake --build build --config release --target libhailort && \
-    cmake --build build --config release --target _pyhailort && \
-    cp build/hailort/libhailort/bindings/python/src/_pyhailort.cpython-${PYTHON_VERSION_NO_DOT}-$(if [ $TARGETARCH == "amd64" ]; then echo 'x86_64'; else echo 'aarch64'; fi )-linux-gnu.so hailort/libhailort/bindings/python/platform/hailo_platform/pyhailort/ && \
-    cp build/hailort/libhailort/src/libhailort.so hailort/libhailort/bindings/python/platform/hailo_platform/pyhailort/
-
-RUN ls -ahl /opt/hailort/build/hailort/libhailort/src/
-RUN ls -ahl /opt/hailort/hailort/libhailort/bindings/python/platform/hailo_platform/pyhailort/
-
-# Remove the existing setup.py if it exists in the target directory
-RUN rm -f /opt/hailort/hailort/libhailort/bindings/python/platform/setup.py
-
-# Copy generate_wheel_conf.py and setup.py
-COPY docker/hailo8l/pyhailort_build_scripts/generate_wheel_conf.py /opt/hailort/hailort/libhailort/bindings/python/platform/generate_wheel_conf.py
-COPY docker/hailo8l/pyhailort_build_scripts/setup.py /opt/hailort/hailort/libhailort/bindings/python/platform/setup.py
-
-# Run the generate_wheel_conf.py script
-RUN python3 /opt/hailort/hailort/libhailort/bindings/python/platform/generate_wheel_conf.py
-
-# Create a wheel file using pip3 wheel
-RUN cd /opt/hailort/hailort/libhailort/bindings/python/platform && \
-    python3 setup.py bdist_wheel --dist-dir /hailo-wheels
+RUN --mount=type=bind,source=docker/hailo8l/install_hailort.sh,target=/deps/install_hailort.sh \
+    /deps/install_hailort.sh
 
 # Use deps as the base image
 FROM deps AS h8l-frigate
 
 # Copy the wheels from the wheels stage
 COPY --from=h8l-wheels /h8l-wheels /deps/h8l-wheels
-COPY --from=build-hailort /hailo-wheels /deps/hailo-wheels
-COPY --from=build-hailort /etc/environment /etc/environment
-RUN CC=$(python3 -c "import sysconfig; import shlex; cc = sysconfig.get_config_var('CC'); cc_cmd = shlex.split(cc)[0]; print(cc_cmd[:-4] if cc_cmd.endswith('-gcc') else cc_cmd)") && \
-    echo "CC=$CC" >> /etc/environment
+COPY --from=hailort /hailo-wheels /deps/hailo-wheels
+COPY --from=hailort /rootfs/ /
 
 # Install the wheels
 RUN pip3 install -U /deps/h8l-wheels/*.whl
 RUN pip3 install -U /deps/hailo-wheels/*.whl
 
-RUN . /etc/environment && \
-    mv /usr/local/lib/python${PYTHON_VERSION}/dist-packages/hailo_platform/pyhailort/libhailort.so /usr/lib/${CC} && \
-    cd /usr/lib/${CC}/ && \
-    ln -s libhailort.so libhailort.so.4.18.0
-
 # Copy base files from the rootfs stage
 COPY --from=rootfs / /
 
-# Set environment variables for Hailo SDK
-ENV PATH="/opt/hailort/bin:${PATH}"
-ENV LD_LIBRARY_PATH="/usr/lib/$(if [ $TARGETARCH == "amd64" ]; then echo 'x86_64'; else echo 'aarch64'; fi )-linux-gnu:${LD_LIBRARY_PATH}"
-
 # Set workdir
 WORKDIR /opt/frigate/

docker/hailo8l/h8l.hcl

@@ -1,3 +1,9 @@
+target wget {
+  dockerfile = "docker/main/Dockerfile"
+  platforms = ["linux/arm64","linux/amd64"]
+  target = "wget"
+}
+
 target wheels {
   dockerfile = "docker/main/Dockerfile"
   platforms = ["linux/arm64","linux/amd64"]
@@ -19,6 +25,7 @@ target rootfs {
 target h8l {
   dockerfile = "docker/hailo8l/Dockerfile"
   contexts = {
+    wget = "target:wget"
     wheels = "target:wheels"
     deps = "target:deps"
     rootfs = "target:rootfs"

docker/hailo8l/install_hailort.sh

@@ -0,0 +1,21 @@
+#!/bin/bash
+
+set -euxo pipefail
+
+hailo_version="4.19.0"
+
+if [[ "${TARGETARCH}" == "amd64" ]]; then
+    arch="x86_64"
+elif [[ "${TARGETARCH}" == "arm64" ]]; then
+    arch="aarch64"
+fi
+
+mkdir -p /rootfs
+
+wget -qO- "https://github.com/frigate-nvr/hailort/releases/download/v${hailo_version}/hailort-${TARGETARCH}.tar.gz" |
+    tar -C /rootfs/ -xzf -
+
+mkdir -p /hailo-wheels
+
+wget -P /hailo-wheels/ "https://github.com/frigate-nvr/hailort/releases/download/v${hailo_version}/hailort-${hailo_version}-cp39-cp39-linux_${arch}.whl"
+

docker/hailo8l/pyhailort_build_scripts/generate_wheel_conf.py

@@ -1,67 +0,0 @@
-import json
-import os
-import platform
-import sys
-import sysconfig
-
-
-def extract_toolchain_info(compiler):
-    # Remove the "-gcc" or "-g++" suffix if present
-    if compiler.endswith("-gcc") or compiler.endswith("-g++"):
-        compiler = compiler.rsplit("-", 1)[0]
-
-    # Extract the toolchain and ABI part (e.g., "gnu")
-    toolchain_parts = compiler.split("-")
-    abi_conventions = next(
-        (part for part in toolchain_parts if part in ["gnu", "musl", "eabi", "uclibc"]),
-        "",
-    )
-
-    return abi_conventions
-
-
-def generate_wheel_conf():
-    conf_file_path = os.path.join(
-        os.path.abspath(os.path.dirname(__file__)), "wheel_conf.json"
-    )
-
-    # Extract current system and Python version information
-    py_version = f"cp{sys.version_info.major}{sys.version_info.minor}"
-    arch = platform.machine()
-    system = platform.system().lower()
-    libc_version = platform.libc_ver()[1]
-
-    # Get the compiler information
-    compiler = sysconfig.get_config_var("CC")
-    abi_conventions = extract_toolchain_info(compiler)
-
-    # Create the new configuration data
-    new_conf_data = {
-        "py_version": py_version,
-        "arch": arch,
-        "system": system,
-        "libc_version": libc_version,
-        "abi": abi_conventions,
-        "extension": {
-            "posix": "so",
-            "nt": "pyd",  # Windows
-        }[os.name],
-    }
-
-    # If the file exists, load the existing data
-    if os.path.isfile(conf_file_path):
-        with open(conf_file_path, "r") as conf_file:
-            conf_data = json.load(conf_file)
-        # Update the existing data with the new data
-        conf_data.update(new_conf_data)
-    else:
-        # If the file does not exist, use the new data
-        conf_data = new_conf_data
-
-    # Write the updated data to the file
-    with open(conf_file_path, "w") as conf_file:
-        json.dump(conf_data, conf_file, indent=4)
-
-
-if __name__ == "__main__":
-    generate_wheel_conf()

docker/hailo8l/pyhailort_build_scripts/setup.py

@@ -1,111 +0,0 @@
-import json
-import os
-
-from setuptools import find_packages, setup
-from wheel.bdist_wheel import bdist_wheel as orig_bdist_wheel
-
-
-class NonPurePythonBDistWheel(orig_bdist_wheel):
-    """Makes the wheel platform-dependent so it can be based on the _pyhailort architecture"""
-
-    def finalize_options(self):
-        orig_bdist_wheel.finalize_options(self)
-        self.root_is_pure = False
-
-
-def _get_hailort_lib_path():
-    lib_filename = "libhailort.so"
-    lib_path = os.path.join(
-        os.path.abspath(os.path.dirname(__file__)),
-        f"hailo_platform/pyhailort/{lib_filename}",
-    )
-    if os.path.exists(lib_path):
-        print(f"Found libhailort shared library at: {lib_path}")
-    else:
-        print(f"Error: libhailort shared library not found at: {lib_path}")
-        raise FileNotFoundError(f"libhailort shared library not found at: {lib_path}")
-    return lib_path
-
-
-def _get_pyhailort_lib_path():
-    conf_file_path = os.path.join(
-        os.path.abspath(os.path.dirname(__file__)), "wheel_conf.json"
-    )
-    if not os.path.isfile(conf_file_path):
-        raise FileNotFoundError(f"Configuration file not found: {conf_file_path}")
-
-    with open(conf_file_path, "r") as conf_file:
-        content = json.load(conf_file)
-        py_version = content["py_version"]
-        arch = content["arch"]
-        system = content["system"]
-        extension = content["extension"]
-        abi = content["abi"]
-
-        # Construct the filename directly
-        lib_filename = f"_pyhailort.cpython-{py_version.split('cp')[1]}-{arch}-{system}-{abi}.{extension}"
-        lib_path = os.path.join(
-            os.path.abspath(os.path.dirname(__file__)),
-            f"hailo_platform/pyhailort/{lib_filename}",
-        )
-
-        if os.path.exists(lib_path):
-            print(f"Found _pyhailort shared library at: {lib_path}")
-        else:
-            print(f"Error: _pyhailort shared library not found at: {lib_path}")
-            raise FileNotFoundError(
-                f"_pyhailort shared library not found at: {lib_path}"
-            )
-
-        return lib_path
-
-
-def _get_package_paths():
-    packages = []
-    pyhailort_lib = _get_pyhailort_lib_path()
-    hailort_lib = _get_hailort_lib_path()
-    if pyhailort_lib:
-        packages.append(pyhailort_lib)
-    if hailort_lib:
-        packages.append(hailort_lib)
-    packages.append(os.path.abspath("hailo_tutorials/notebooks/*"))
-    packages.append(os.path.abspath("hailo_tutorials/hefs/*"))
-    return packages
-
-
-if __name__ == "__main__":
-    setup(
-        author="Hailo team",
-        author_email="contact@hailo.ai",
-        cmdclass={
-            "bdist_wheel": NonPurePythonBDistWheel,
-        },
-        description="HailoRT",
-        entry_points={
-            "console_scripts": [
-                "hailo=hailo_platform.tools.hailocli.main:main",
-            ]
-        },
-        install_requires=[
-            "argcomplete",
-            "contextlib2",
-            "future",
-            "netaddr",
-            "netifaces",
-            "verboselogs",
-            "numpy==1.23.3",
-        ],
-        name="hailort",
-        package_data={
-            "hailo_platform": _get_package_paths(),
-        },
-        packages=find_packages(),
-        platforms=[
-            "linux_x86_64",
-            "linux_aarch64",
-            "win_amd64",
-        ],
-        url="https://hailo.ai/",
-        version="4.17.0",
-        zip_safe=False,
-    )

docker/hailo8l/user_installation.sh

@@ -13,7 +13,7 @@ else
 fi
 
 # Clone the HailoRT driver repository
-git clone --depth 1 --branch v4.18.0 https://github.com/hailo-ai/hailort-drivers.git
+git clone --depth 1 --branch v4.19.0 https://github.com/hailo-ai/hailort-drivers.git
 
 # Build and install the HailoRT driver
 cd hailort-drivers/linux/pcie
@@ -38,7 +38,7 @@ cd ../../
 if [ ! -d /lib/firmware/hailo ]; then
   sudo mkdir /lib/firmware/hailo
 fi
-sudo mv hailo8_fw.4.17.0.bin /lib/firmware/hailo/hailo8_fw.bin
+sudo mv hailo8_fw.*.bin /lib/firmware/hailo/hailo8_fw.bin
 
 # Install udev rules
 sudo cp ./linux/pcie/51-hailo-udev.rules /etc/udev/rules.d/

docker/main/Dockerfile

@@ -180,9 +180,6 @@ RUN /build_pysqlite3.sh
 COPY docker/main/requirements-wheels.txt /requirements-wheels.txt
 RUN pip3 wheel --wheel-dir=/wheels -r /requirements-wheels.txt
 
-COPY docker/main/requirements-wheels-post.txt /requirements-wheels-post.txt
-RUN pip3 wheel --no-deps --wheel-dir=/wheels-post -r /requirements-wheels-post.txt
-
 
 # Collect deps in a single layer
 FROM scratch AS deps-rootfs
@@ -214,6 +211,9 @@ ENV TOKENIZERS_PARALLELISM=true
 # https://github.com/huggingface/transformers/issues/27214
 ENV TRANSFORMERS_NO_ADVISORY_WARNINGS=1
 
+# Set OpenCV ffmpeg loglevel to fatal: https://ffmpeg.org/doxygen/trunk/log_8h.html
+ENV OPENCV_FFMPEG_LOGLEVEL=8
+
 ENV PATH="/usr/local/go2rtc/bin:/usr/local/tempio/bin:/usr/local/nginx/sbin:${PATH}"
 ENV LIBAVFORMAT_VERSION_MAJOR=60
 
@@ -225,14 +225,6 @@ RUN --mount=type=bind,from=wheels,source=/wheels,target=/deps/wheels \
     python3 -m pip install --upgrade pip && \
     pip3 install -U /deps/wheels/*.whl
 
-# We have to uninstall this dependency specifically
-# as it will break onnxruntime-openvino
-RUN pip3 uninstall -y onnxruntime
-
-RUN --mount=type=bind,from=wheels,source=/wheels-post,target=/deps/wheels \
-    python3 -m pip install --upgrade pip && \
-    pip3 install -U /deps/wheels/*.whl
-
 COPY --from=deps-rootfs / /
 
 RUN ldconfig

docker/main/install_deps.sh

@@ -8,6 +8,7 @@ apt-get -qq install --no-install-recommends -y \
     apt-transport-https \
     gnupg \
     wget \
+    lbzip2 \
     procps vainfo \
     unzip locales tzdata libxml2 xz-utils \
     python3.9 \
@@ -45,7 +46,7 @@ if [[ "${TARGETARCH}" == "amd64" ]]; then
     wget -qO btbn-ffmpeg.tar.xz "https://github.com/NickM-27/FFmpeg-Builds/releases/download/autobuild-2022-07-31-12-37/ffmpeg-n5.1-2-g915ef932a3-linux64-gpl-5.1.tar.xz"
     tar -xf btbn-ffmpeg.tar.xz -C /usr/lib/ffmpeg/5.0 --strip-components 1
     rm -rf btbn-ffmpeg.tar.xz /usr/lib/ffmpeg/5.0/doc /usr/lib/ffmpeg/5.0/bin/ffplay
-    wget -qO btbn-ffmpeg.tar.xz "https://github.com/BtbN/FFmpeg-Builds/releases/download/autobuild-2024-09-30-15-36/ffmpeg-n7.1-linux64-gpl-7.1.tar.xz"
+    wget -qO btbn-ffmpeg.tar.xz "https://github.com/NickM-27/FFmpeg-Builds/releases/download/autobuild-2024-09-19-12-51/ffmpeg-n7.0.2-18-g3e6cec1286-linux64-gpl-7.0.tar.xz"
     tar -xf btbn-ffmpeg.tar.xz -C /usr/lib/ffmpeg/7.0 --strip-components 1
     rm -rf btbn-ffmpeg.tar.xz /usr/lib/ffmpeg/7.0/doc /usr/lib/ffmpeg/7.0/bin/ffplay
 fi
@@ -57,7 +58,7 @@ if [[ "${TARGETARCH}" == "arm64" ]]; then
     wget -qO btbn-ffmpeg.tar.xz "https://github.com/NickM-27/FFmpeg-Builds/releases/download/autobuild-2022-07-31-12-37/ffmpeg-n5.1-2-g915ef932a3-linuxarm64-gpl-5.1.tar.xz"
     tar -xf btbn-ffmpeg.tar.xz -C /usr/lib/ffmpeg/5.0 --strip-components 1
     rm -rf btbn-ffmpeg.tar.xz /usr/lib/ffmpeg/5.0/doc /usr/lib/ffmpeg/5.0/bin/ffplay
-    wget -qO btbn-ffmpeg.tar.xz "https://github.com/BtbN/FFmpeg-Builds/releases/download/autobuild-2024-09-30-15-36/ffmpeg-n7.1-linuxarm64-gpl-7.1.tar.xz"
+    wget -qO btbn-ffmpeg.tar.xz "https://github.com/NickM-27/FFmpeg-Builds/releases/download/autobuild-2024-09-19-12-51/ffmpeg-n7.0.2-18-g3e6cec1286-linuxarm64-gpl-7.0.tar.xz"
     tar -xf btbn-ffmpeg.tar.xz -C /usr/lib/ffmpeg/7.0 --strip-components 1
     rm -rf btbn-ffmpeg.tar.xz /usr/lib/ffmpeg/7.0/doc /usr/lib/ffmpeg/7.0/bin/ffplay
 fi
@@ -76,6 +77,9 @@ if [[ "${TARGETARCH}" == "amd64" ]]; then
     apt-get -qq install --no-install-recommends --no-install-suggests -y \
         i965-va-driver-shaders
 
+    # intel packages use zst compression so we need to update dpkg
+    apt-get install -y dpkg
+
     rm -f /etc/apt/sources.list.d/debian-bookworm.list
 
     # use intel apt intel packages
@@ -83,8 +87,8 @@ if [[ "${TARGETARCH}" == "amd64" ]]; then
     echo "deb [arch=amd64 signed-by=/usr/share/keyrings/intel-graphics.gpg] https://repositories.intel.com/gpu/ubuntu jammy client" | tee /etc/apt/sources.list.d/intel-gpu-jammy.list
     apt-get -qq update
     apt-get -qq install --no-install-recommends --no-install-suggests -y \
-        intel-opencl-icd intel-level-zero-gpu intel-media-va-driver-non-free \
-        libmfx1 libmfxgen1 libvpl2
+        intel-opencl-icd=24.35.30872.31-996~22.04 intel-level-zero-gpu=1.3.29735.27-914~22.04 intel-media-va-driver-non-free=24.3.3-996~22.04 \
+        libmfx1=23.2.2-880~22.04 libmfxgen1=24.2.4-914~22.04 libvpl2=1:2.13.0.0-996~22.04
 
     rm -f /usr/share/keyrings/intel-graphics.gpg
     rm -f /etc/apt/sources.list.d/intel-gpu-jammy.list

docker/main/requirements-wheels-post.txt

@@ -1,3 +0,0 @@
-# ONNX
-onnxruntime-openvino == 1.19.* ; platform_machine == 'x86_64'
-onnxruntime == 1.19.* ; platform_machine == 'aarch64'

docker/main/requirements-wheels.txt

@@ -1,9 +1,11 @@
 click == 8.1.*
 # FastAPI
+aiohttp == 3.11.2
+starlette == 0.41.2
 starlette-context == 0.3.6
-fastapi == 0.115.0
+fastapi == 0.115.*
 uvicorn == 0.30.*
-slowapi == 0.1.9
+slowapi == 0.1.*
 imutils == 0.5.*
 joserfc == 1.0.*
 pathvalidate == 3.2.*
@@ -16,10 +18,10 @@ paho-mqtt == 2.1.*
 pandas == 2.2.*
 peewee == 3.17.*
 peewee_migrate == 1.13.*
-psutil == 5.9.*
+psutil == 6.1.*
 pydantic == 2.8.*
 git+https://github.com/fbcotter/py3nvml#egg=py3nvml
-pytz == 2024.1
+pytz == 2024.*
 pyzmq == 26.2.*
 ruamel.yaml == 0.18.*
 tzlocal == 5.2
@@ -30,11 +32,12 @@ norfair == 2.2.*
 setproctitle == 1.3.*
 ws4py == 0.5.*
 unidecode == 1.3.*
-# OpenVino (ONNX installed in wheels-post)
+# OpenVino & ONNX
 openvino == 2024.3.*
+onnxruntime-openvino == 1.19.* ; platform_machine == 'x86_64'
+onnxruntime == 1.19.* ; platform_machine == 'aarch64'
 # Embeddings
 transformers == 4.45.*
-onnx_clip == 4.0.*
 # Generative AI
 google-generativeai == 0.8.*
 ollama == 0.3.*

docker/main/rootfs/usr/local/go2rtc/create_config.py

@@ -165,7 +165,7 @@ if config.get("birdseye", {}).get("restream", False):
     birdseye: dict[str, any] = config.get("birdseye")
 
     input = f"-f rawvideo -pix_fmt yuv420p -video_size {birdseye.get('width', 1280)}x{birdseye.get('height', 720)} -r 10 -i {BIRDSEYE_PIPE}"
-    ffmpeg_cmd = f"exec:{parse_preset_hardware_acceleration_encode(ffmpeg_path, config.get('ffmpeg', {}).get('hwaccel_args'), input, '-rtsp_transport tcp -f rtsp {output}')}"
+    ffmpeg_cmd = f"exec:{parse_preset_hardware_acceleration_encode(ffmpeg_path, config.get('ffmpeg', {}).get('hwaccel_args', ''), input, '-rtsp_transport tcp -f rtsp {output}')}"
 
     if go2rtc_config.get("streams"):
         go2rtc_config["streams"]["birdseye"] = ffmpeg_cmd

docker/tensorrt/Dockerfile.arm64

@@ -10,8 +10,8 @@ ARG DEBIAN_FRONTEND
 # Use a separate container to build wheels to prevent build dependencies in final image
 RUN apt-get -qq update \
     && apt-get -qq install -y --no-install-recommends \
-       python3.9 python3.9-dev \
-       wget build-essential cmake git \
+    python3.9 python3.9-dev \
+    wget build-essential cmake git \
     && rm -rf /var/lib/apt/lists/*
 
 # Ensure python3 defaults to python3.9
@@ -41,7 +41,11 @@ RUN --mount=type=bind,source=docker/tensorrt/detector/build_python_tensorrt.sh,t
     && TENSORRT_VER=$(cat /etc/TENSORRT_VER) /deps/build_python_tensorrt.sh
 
 COPY docker/tensorrt/requirements-arm64.txt /requirements-tensorrt.txt
-RUN pip3 wheel --wheel-dir=/trt-wheels -r /requirements-tensorrt.txt
+ADD https://nvidia.box.com/shared/static/9aemm4grzbbkfaesg5l7fplgjtmswhj8.whl /tmp/onnxruntime_gpu-1.15.1-cp39-cp39-linux_aarch64.whl
+
+RUN pip3 uninstall -y onnxruntime-openvino \
+    && pip3 wheel --wheel-dir=/trt-wheels -r /requirements-tensorrt.txt \
+    && pip3 install --no-deps /tmp/onnxruntime_gpu-1.15.1-cp39-cp39-linux_aarch64.whl
 
 FROM build-wheels AS trt-model-wheels
 ARG DEBIAN_FRONTEND

docker/tensorrt/Dockerfile.base

@@ -25,7 +25,7 @@ ENV S6_CMD_WAIT_FOR_SERVICES_MAXTIME=0
 COPY --from=trt-deps /usr/local/lib/libyolo_layer.so /usr/local/lib/libyolo_layer.so
 COPY --from=trt-deps /usr/local/src/tensorrt_demos /usr/local/src/tensorrt_demos
 COPY docker/tensorrt/detector/rootfs/ /
-ENV YOLO_MODELS="yolov7-320"
+ENV YOLO_MODELS=""
 
 HEALTHCHECK --start-period=600s --start-interval=5s --interval=15s --timeout=5s --retries=3 \
     CMD curl --fail --silent --show-error http://127.0.0.1:5000/api/version || exit 1

docker/tensorrt/detector/rootfs/etc/s6-overlay/s6-rc.d/trt-model-prepare/run

@@ -11,6 +11,7 @@ set -o errexit -o nounset -o pipefail
 MODEL_CACHE_DIR=${MODEL_CACHE_DIR:-"/config/model_cache/tensorrt"}
 TRT_VER=${TRT_VER:-$(cat /etc/TENSORRT_VER)}
 OUTPUT_FOLDER="${MODEL_CACHE_DIR}/${TRT_VER}"
+YOLO_MODELS=${YOLO_MODELS:-""}
 
 # Create output folder
 mkdir -p ${OUTPUT_FOLDER}
@@ -19,6 +20,11 @@ FIRST_MODEL=true
 MODEL_DOWNLOAD=""
 MODEL_CONVERT=""
 
+if [ -z "$YOLO_MODELS"]; then
+    echo "tensorrt model preparation disabled"
+    exit 0
+fi
+
 for model in ${YOLO_MODELS//,/ }
 do
     # Remove old link in case path/version changed

docker/tensorrt/requirements-amd64.txt

@@ -9,6 +9,6 @@ nvidia-cuda-runtime-cu11 == 11.8.*; platform_machine == 'x86_64'
 nvidia-cublas-cu11 == 11.11.3.6; platform_machine == 'x86_64'
 nvidia-cudnn-cu11 == 8.6.0.*; platform_machine == 'x86_64'
 nvidia-cufft-cu11==10.*; platform_machine == 'x86_64'
-onnx==1.14.0; platform_machine == 'x86_64'
-onnxruntime-gpu==1.17.*; platform_machine == 'x86_64'
+onnx==1.16.*; platform_machine == 'x86_64'
+onnxruntime-gpu==1.18.*; platform_machine == 'x86_64'
 protobuf==3.20.3; platform_machine == 'x86_64'

docker/tensorrt/requirements-arm64.txt

@@ -1 +1 @@
-cuda-python == 11.7; platform_machine == 'aarch64'
+cuda-python == 11.7; platform_machine == 'aarch64'

docs/docs/configuration/camera_specific.md

@@ -181,7 +181,7 @@ go2rtc:
       - rtspx://192.168.1.1:7441/abcdefghijk
 ```
 
-[See the go2rtc docs for more information](https://github.com/AlexxIT/go2rtc/tree/v1.9.4#source-rtsp)
+[See the go2rtc docs for more information](https://github.com/AlexxIT/go2rtc/tree/v1.9.2#source-rtsp)
 
 In the Unifi 2.0 update Unifi Protect Cameras had a change in audio sample rate which causes issues for ffmpeg. The input rate needs to be set for record if used directly with unifi protect.
 

docs/docs/configuration/cameras.md

@@ -109,7 +109,7 @@ This list of working and non-working PTZ cameras is based on user feedback.
 | Reolink E1 Zoom              |      ✅      |      ❌      |                                                                                                                                                 |
 | Reolink RLC-823A 16x         |      ✅      |      ❌      |                                                                                                                                                 |
 | Speco O8P32X                 |      ✅      |      ❌      |                                                                                                                                                 |
-| Sunba 405-D20X               |      ✅      |      ❌      |                                                                                                                                                 |
+| Sunba 405-D20X               |      ✅      |      ❌      | Incomplete ONVIF support reported on original, and 4k models. All models are suspected incompatable.                                            |
 | Tapo                         |      ✅      |      ❌      | Many models supported, ONVIF Service Port: 2020                                                                                                 |
 | Uniview IPC672LR-AX4DUPK     |      ✅      |      ❌      | Firmware says FOV relative movement is supported, but camera doesn't actually move when sending ONVIF commands                                  |
 | Uniview IPC6612SR-X33-VG     |      ✅      |      ✅      | Leave `calibrate_on_startup` as `False`. A user has reported that zooming with `absolute` is working.                                           |

docs/docs/configuration/genai.md

@@ -3,9 +3,13 @@ id: genai
 title: Generative AI
 ---
 
-Generative AI can be used to automatically generate descriptions based on the thumbnails of your tracked objects. This helps with [Semantic Search](/configuration/semantic_search) in Frigate by providing detailed text descriptions as a basis of the search query.
+Generative AI can be used to automatically generate descriptive text based on the thumbnails of your tracked objects. This helps with [Semantic Search](/configuration/semantic_search) in Frigate to provide more context about your tracked objects. Descriptions are accessed via the _Explore_ view in the Frigate UI by clicking on a tracked object's thumbnail.
 
-Semantic Search must be enabled to use Generative AI. Descriptions are accessed via the _Explore_ view in the Frigate UI by clicking on a tracked object's thumbnail.
+:::info
+
+Semantic Search must be enabled to use Generative AI.
+
+:::
 
 ## Configuration
 
@@ -29,11 +33,21 @@ cameras:
 
 ## Ollama
 
-[Ollama](https://ollama.com/) allows you to self-host large language models and keep everything running locally. It provides a nice API over [llama.cpp](https://github.com/ggerganov/llama.cpp). It is highly recommended to host this server on a machine with an Nvidia graphics card, or on a Apple silicon Mac for best performance. Most of the 7b parameter 4-bit vision models will fit inside 8GB of VRAM. There is also a [docker container](https://hub.docker.com/r/ollama/ollama) available.
+:::warning
+
+Using Ollama on CPU is not recommended, high inference times make using Generative AI impractical.
+
+:::
+
+[Ollama](https://ollama.com/) allows you to self-host large language models and keep everything running locally. It provides a nice API over [llama.cpp](https://github.com/ggerganov/llama.cpp). It is highly recommended to host this server on a machine with an Nvidia graphics card, or on a Apple silicon Mac for best performance.
+
+Most of the 7b parameter 4-bit vision models will fit inside 8GB of VRAM. There is also a [Docker container](https://hub.docker.com/r/ollama/ollama) available.
+
+Parallel requests also come with some caveats. You will need to set `OLLAMA_NUM_PARALLEL=1` and choose a `OLLAMA_MAX_QUEUE` and `OLLAMA_MAX_LOADED_MODELS` values that are appropriate for your hardware and preferences. See the [Ollama documentation](https://github.com/ollama/ollama/blob/main/docs/faq.md#how-does-ollama-handle-concurrent-requests).
 
 ### Supported Models
 
-You must use a vision capable model with Frigate. Current model variants can be found [in their model library](https://ollama.com/library). At the time of writing, this includes `llava`, `llava-llama3`, `llava-phi3`, and `moondream`.
+You must use a vision capable model with Frigate. Current model variants can be found [in their model library](https://ollama.com/library). At the time of writing, this includes `llava`, `llava-llama3`, `llava-phi3`, and `moondream`. Note that Frigate will not automatically download the model you specify in your config, you must download the model to your local instance of Ollama first i.e. by running `ollama pull llava:7b` on your Ollama server/Docker container. Note that the model specified in Frigate's config must match the downloaded model tag.
 
 :::note
 
@@ -48,7 +62,7 @@ genai:
   enabled: True
   provider: ollama
   base_url: http://localhost:11434
-  model: llava
+  model: llava:7b
 ```
 
 ## Google Gemini
@@ -122,12 +136,22 @@ genai:
   api_key: "{FRIGATE_OPENAI_API_KEY}"
 ```
 
+## Usage and Best Practices
+
+Frigate's thumbnail search excels at identifying specific details about tracked objects – for example, using an "image caption" approach to find a "person wearing a yellow vest," "a white dog running across the lawn," or "a red car on a residential street." To enhance this further, Frigate’s default prompts are designed to ask your AI provider about the intent behind the object's actions, rather than just describing its appearance.
+
+While generating simple descriptions of detected objects is useful, understanding intent provides a deeper layer of insight. Instead of just recognizing "what" is in a scene, Frigate’s default prompts aim to infer "why" it might be there or "what" it could do next. Descriptions tell you what’s happening, but intent gives context. For instance, a person walking toward a door might seem like a visitor, but if they’re moving quickly after hours, you can infer a potential break-in attempt. Detecting a person loitering near a door at night can trigger an alert sooner than simply noting "a person standing by the door," helping you respond based on the situation’s context.
+
+### Using GenAI for notifications
+
+Frigate provides an [MQTT topic](/integrations/mqtt), `frigate/tracked_object_update`, that is updated with a JSON payload containing `event_id` and `description` when your AI provider returns a description for a tracked object. This description could be used directly in notifications, such as sending alerts to your phone or making audio announcements. If additional details from the tracked object are needed, you can query the [HTTP API](/integrations/api/event-events-event-id-get) using the `event_id`, eg: `http://frigate_ip:5000/api/events/<event_id>`.
+
 ## Custom Prompts
 
 Frigate sends multiple frames from the tracked object along with a prompt to your Generative AI provider asking it to generate a description. The default prompt is as follows:
 
 ```
-Describe the {label} in the sequence of images with as much detail as possible. Do not describe the background.
+Analyze the sequence of images containing the {label}. Focus on the likely intent or behavior of the {label} based on its actions and movement, rather than describing its appearance or the surroundings. Consider what the {label} is doing, why, and what it might do next.
 ```
 
 :::tip
@@ -144,25 +168,25 @@ genai:
   provider: ollama
   base_url: http://localhost:11434
   model: llava
-  prompt: "Describe the {label} in these images from the {camera} security camera."
+  prompt: "Analyze the {label} in these images from the {camera} security camera. Focus on the actions, behavior, and potential intent of the {label}, rather than just describing its appearance."
   object_prompts:
-    person: "Describe the main person in these images (gender, age, clothing, activity, etc). Do not include where the activity is occurring (sidewalk, concrete, driveway, etc)."
-    car: "Label the primary vehicle in these images with just the name of the company if it is a delivery vehicle, or the color make and model."
+    person: "Examine the main person in these images. What are they doing and what might their actions suggest about their intent (e.g., approaching a door, leaving an area, standing still)? Do not describe the surroundings or static details."
+    car: "Observe the primary vehicle in these images. Focus on its movement, direction, or purpose (e.g., parking, approaching, circling). If it's a delivery vehicle, mention the company."
 ```
 
 Prompts can also be overriden at the camera level to provide a more detailed prompt to the model about your specific camera, if you desire. By default, descriptions will be generated for all tracked objects and all zones. But you can also optionally specify `objects` and `required_zones` to only generate descriptions for certain tracked objects or zones.
 
-Optionally, you can generate the description using a snapshot (if enabled) by setting `use_snapshot` to `True`. By default, this is set to `False`, which sends the thumbnails collected over the object's lifetime to the model. Using a snapshot provides the AI with a higher-resolution image (typically downscaled by the AI itself), but the trade-off is that only a single image is used, which might limit the model's ability to determine object movement or direction.
+Optionally, you can generate the description using a snapshot (if enabled) by setting `use_snapshot` to `True`. By default, this is set to `False`, which sends the uncompressed images from the `detect` stream collected over the object's lifetime to the model. Once the object lifecycle ends, only a single compressed and cropped thumbnail is saved with the tracked object. Using a snapshot might be useful when you want to _regenerate_ a tracked object's description as it will provide the AI with a higher-quality image (typically downscaled by the AI itself) than the cropped/compressed thumbnail. Using a snapshot otherwise has a trade-off in that only a single image is sent to your provider, which will limit the model's ability to determine object movement or direction.
 
 ```yaml
 cameras:
   front_door:
     genai:
       use_snapshot: True
-      prompt: "Describe the {label} in these images from the {camera} security camera at the front door of a house, aimed outward toward the street."
+      prompt: "Analyze the {label} in these images from the {camera} security camera at the front door. Focus on the actions and potential intent of the {label}."
       object_prompts:
-        person: "Describe the main person in these images (gender, age, clothing, activity, etc). Do not include where the activity is occurring (sidewalk, concrete, driveway, etc). If delivering a package, include the company the package is from."
-        cat: "Describe the cat in these images (color, size, tail). Indicate whether or not the cat is by the flower pots. If the cat is chasing a mouse, make up a name for the mouse."
+        person: "Examine the person in these images. What are they doing, and how might their actions suggest their purpose (e.g., delivering something, approaching, leaving)? If they are carrying or interacting with a package, include details about its source or destination."
+        cat: "Observe the cat in these images. Focus on its movement and intent (e.g., wandering, hunting, interacting with objects). If the cat is near the flower pots or engaging in any specific actions, mention it."
       objects:
         - person
         - cat

docs/docs/configuration/motion_detection.md

@@ -92,10 +92,16 @@ motion:
   lightning_threshold: 0.8
 ```
 
-:::tip
+:::warning
 
 Some cameras like doorbell cameras may have missed detections when someone walks directly in front of the camera and the lightning_threshold causes motion detection to be re-calibrated. In this case, it may be desirable to increase the `lightning_threshold` to ensure these objects are not missed.
 
 :::
 
+:::note
+
+Lightning threshold does not stop motion based recordings from being saved.
+
+:::
+
 Large changes in motion like PTZ moves and camera switches between Color and IR mode should result in no motion detection. This is done via the `lightning_threshold` configuration. It is defined as the percentage of the image used to detect lightning or other substantial changes where motion detection needs to recalibrate. Increasing this value will make motion detection more likely to consider lightning or IR mode changes as valid motion. Decreasing this value will make motion detection more likely to ignore large amounts of motion such as a person approaching a doorbell camera.

docs/docs/configuration/object_detectors.md

@@ -22,14 +22,14 @@ Frigate supports multiple different detectors that work on different types of ha
 - [ONNX](#onnx): OpenVINO will automatically be detected and used as a detector in the default Frigate image when a supported ONNX model is configured.
 
 **Nvidia**
-- [TensortRT](#nvidia-tensorrt-detector): TensorRT can run on Nvidia GPUs, using one of many default models.
-- [ONNX](#onnx): TensorRT will automatically be detected and used as a detector in the `-tensorrt` Frigate image when a supported ONNX model is configured.
+- [TensortRT](#nvidia-tensorrt-detector): TensorRT can run on Nvidia GPUs and Jetson devices, using one of many default models.
+- [ONNX](#onnx): TensorRT will automatically be detected and used as a detector in the `-tensorrt` or `-tensorrt-jp(4/5)` Frigate images when a supported ONNX model is configured.
 
 **Rockchip**
 - [RKNN](#rockchip-platform): RKNN models can run on Rockchip devices with included NPUs.
 
 **For Testing**
-- [CPU Detector (not recommended for actual use](#cpu-detector-not-recommended): Use a CPU to run tflite model, this is not recommended and in most cases OpenVINO can be used in CPU mode with better results. 
+- [CPU Detector (not recommended for actual use](#cpu-detector-not-recommended): Use a CPU to run tflite model, this is not recommended and in most cases OpenVINO can be used in CPU mode with better results.
 
 :::
 
@@ -223,7 +223,7 @@ The model used for TensorRT must be preprocessed on the same hardware platform t
 
 The Frigate image will generate model files during startup if the specified model is not found. Processed models are stored in the `/config/model_cache` folder. Typically the `/config` path is mapped to a directory on the host already and the `model_cache` does not need to be mapped separately unless the user wants to store it in a different location on the host.
 
-By default, the `yolov7-320` model will be generated, but this can be overridden by specifying the `YOLO_MODELS` environment variable in Docker. One or more models may be listed in a comma-separated format, and each one will be generated. To select no model generation, set the variable to an empty string, `YOLO_MODELS=""`. Models will only be generated if the corresponding `{model}.trt` file is not present in the `model_cache` folder, so you can force a model to be regenerated by deleting it from your Frigate data folder.
+By default, no models will be generated, but this can be overridden by specifying the `YOLO_MODELS` environment variable in Docker. One or more models may be listed in a comma-separated format, and each one will be generated. Models will only be generated if the corresponding `{model}.trt` file is not present in the `model_cache` folder, so you can force a model to be regenerated by deleting it from your Frigate data folder.
 
 If you have a Jetson device with DLAs (Xavier or Orin), you can generate a model that will run on the DLA by appending `-dla` to your model name, e.g. specify `YOLO_MODELS=yolov7-320-dla`. The model will run on DLA0 (Frigate does not currently support DLA1). DLA-incompatible layers will fall back to running on the GPU.
 
@@ -264,7 +264,7 @@ An example `docker-compose.yml` fragment that converts the `yolov4-608` and `yol
 ```yml
 frigate:
   environment:
-    - YOLO_MODELS=yolov4-608,yolov7x-640
+    - YOLO_MODELS=yolov7-320,yolov7x-640
     - USE_FP16=false
 ```
 
@@ -415,6 +415,24 @@ Note that the labelmap uses a subset of the complete COCO label set that has onl
 
 ONNX is an open format for building machine learning models, Frigate supports running ONNX models on CPU, OpenVINO, and TensorRT. On startup Frigate will automatically try to use a GPU if one is available.
 
+:::info
+
+If the correct build is used for your GPU then the GPU will be detected and used automatically.
+
+- **AMD**
+
+  - ROCm will automatically be detected and used with the ONNX detector in the `-rocm` Frigate image.
+
+- **Intel**
+
+  - OpenVINO will automatically be detected and used with the ONNX detector in the default Frigate image.
+
+- **Nvidia**
+  - Nvidia GPUs will automatically be detected and used with the ONNX detector in the `-tensorrt` Frigate image.
+  - Jetson devices will automatically be detected and used with the ONNX detector in the `-tensorrt-jp(4/5)` Frigate image.
+
+:::
+
 :::tip
 
 When using many cameras one detector may not be enough to keep up. Multiple detectors can be defined assuming GPU resources are available. An example configuration would be:
@@ -457,6 +475,7 @@ model:
   width: 320 # <--- should match whatever was set in notebook
   height: 320 # <--- should match whatever was set in notebook
   input_pixel_format: bgr
+  input_tensor: nchw
   path: /config/yolo_nas_s.onnx
   labelmap_path: /labelmap/coco-80.txt
 ```

docs/docs/configuration/objects.md

@@ -5,7 +5,7 @@ title: Available Objects
 
 import labels from "../../../labelmap.txt";
 
-Frigate includes the object models listed below from the Google Coral test data.
+Frigate includes the object labels listed below from the Google Coral test data.
 
 Please note:
 

docs/docs/configuration/reference.md

@@ -518,6 +518,9 @@ semantic_search:
   enabled: False
   # Optional: Re-index embeddings database from historical tracked objects (default: shown below)
   reindex: False
+  # Optional: Set the model size used for embeddings. (default: shown below)
+  # NOTE: small model runs on CPU and large model runs on GPU
+  model_size: "small"
 
 # Optional: Configuration for AI generated tracked object descriptions
 # NOTE: Semantic Search must be enabled for this to do anything.
@@ -545,10 +548,12 @@ genai:
 # Uses https://github.com/AlexxIT/go2rtc (v1.9.2)
 go2rtc:
 
-# Optional: jsmpeg stream configuration for WebUI
+# Optional: Live stream configuration for WebUI.
+# NOTE: Can be overridden at the camera level
 live:
-  # Optional: Set the name of the stream that should be used for live view
-  # in frigate WebUI. (default: name of camera)
+  # Optional: Set the name of the stream configured in go2rtc
+  # that should be used for live view in frigate WebUI. (default: name of camera)
+  # NOTE: In most cases this should be set at the camera level only.
   stream_name: camera_name
   # Optional: Set the height of the jsmpeg stream. (default: 720)
   # This must be less than or equal to the height of the detect stream. Lower resolutions

docs/docs/configuration/restream.md

@@ -7,7 +7,7 @@ title: Restream
 
 Frigate can restream your video feed as an RTSP feed for other applications such as Home Assistant to utilize it at `rtsp://<frigate_host>:8554/<camera_name>`. Port 8554 must be open. [This allows you to use a video feed for detection in Frigate and Home Assistant live view at the same time without having to make two separate connections to the camera](#reduce-connections-to-camera). The video feed is copied from the original video feed directly to avoid re-encoding. This feed does not include any annotation by Frigate.
 
-Frigate uses [go2rtc](https://github.com/AlexxIT/go2rtc/tree/v1.9.4) to provide its restream and MSE/WebRTC capabilities. The go2rtc config is hosted at the `go2rtc` in the config, see [go2rtc docs](https://github.com/AlexxIT/go2rtc/tree/v1.9.4#configuration) for more advanced configurations and features.
+Frigate uses [go2rtc](https://github.com/AlexxIT/go2rtc/tree/v1.9.2) to provide its restream and MSE/WebRTC capabilities. The go2rtc config is hosted at the `go2rtc` in the config, see [go2rtc docs](https://github.com/AlexxIT/go2rtc/tree/v1.9.2#configuration) for more advanced configurations and features.
 
 :::note
 
@@ -134,7 +134,7 @@ cameras:
 
 ## Advanced Restream Configurations
 
-The [exec](https://github.com/AlexxIT/go2rtc/tree/v1.9.4#source-exec) source in go2rtc can be used for custom ffmpeg commands. An example is below:
+The [exec](https://github.com/AlexxIT/go2rtc/tree/v1.9.2#source-exec) source in go2rtc can be used for custom ffmpeg commands. An example is below:
 
 NOTE: The output will need to be passed with two curly braces `{{output}}`
 

docs/docs/configuration/semantic_search.md

@@ -5,13 +5,21 @@ title: Using Semantic Search
 
 Semantic Search in Frigate allows you to find tracked objects within your review items using either the image itself, a user-defined text description, or an automatically generated one. This feature works by creating _embeddings_ — numerical vector representations — for both the images and text descriptions of your tracked objects. By comparing these embeddings, Frigate assesses their similarities to deliver relevant search results.
 
-Frigate has support for two models to create embeddings, both of which run locally: [OpenAI CLIP](https://openai.com/research/clip) and [all-MiniLM-L6-v2](https://huggingface.co/sentence-transformers/all-MiniLM-L6-v2). Embeddings are then saved to Frigate's database.
+Frigate has support for [Jina AI's CLIP model](https://huggingface.co/jinaai/jina-clip-v1) to create embeddings, which runs locally. Embeddings are then saved to Frigate's database.
 
 Semantic Search is accessed via the _Explore_ view in the Frigate UI.
 
+## Minimum System Requirements
+
+Semantic Search works by running a large AI model locally on your system. Small or underpowered systems like a Raspberry Pi will not run Semantic Search reliably or at all.
+
+A minimum of 8GB of RAM is required to use Semantic Search. A GPU is not strictly required but will provide a significant performance increase over CPU-only systems.
+
+For best performance, 16GB or more of RAM and a dedicated GPU are recommended.
+
 ## Configuration
 
-Semantic search is disabled by default, and must be enabled in your config file before it can be used. Semantic Search is a global configuration setting.
+Semantic Search is disabled by default, and must be enabled in your config file before it can be used. Semantic Search is a global configuration setting.
 
 ```yaml
 semantic_search:
@@ -27,18 +35,58 @@ If you are enabling the Search feature for the first time, be advised that Friga
 
 :::
 
-### OpenAI CLIP
+### Jina AI CLIP
 
-This model is able to embed both images and text into the same vector space, which allows `image -> image` and `text -> image` similarity searches. Frigate uses this model on tracked objects to encode the thumbnail image and store it in the database. When searching for tracked objects via text in the search box, Frigate will perform a `text -> image` similarity search against this embedding. When clicking "Find Similar" in the tracked object detail pane, Frigate will perform an `image -> image` similarity search to retrieve the closest matching thumbnails.
+The vision model is able to embed both images and text into the same vector space, which allows `image -> image` and `text -> image` similarity searches. Frigate uses this model on tracked objects to encode the thumbnail image and store it in the database. When searching for tracked objects via text in the search box, Frigate will perform a `text -> image` similarity search against this embedding. When clicking "Find Similar" in the tracked object detail pane, Frigate will perform an `image -> image` similarity search to retrieve the closest matching thumbnails.
 
-### all-MiniLM-L6-v2
+The text model is used to embed tracked object descriptions and perform searches against them. Descriptions can be created, viewed, and modified on the Search page when clicking on the gray tracked object chip at the top left of each review item. See [the Generative AI docs](/configuration/genai.md) for more information on how to automatically generate tracked object descriptions.
 
-This is a sentence embedding model that has been fine tuned on over 1 billion sentence pairs. This model is used to embed tracked object descriptions and perform searches against them. Descriptions can be created, viewed, and modified on the Search page when clicking on the gray tracked object chip at the top left of each review item. See [the Generative AI docs](/configuration/genai.md) for more information on how to automatically generate tracked object descriptions.
+Differently weighted CLIP models are available and can be selected by setting the `model_size` config option as `small` or `large`:
 
-## Usage
+```yaml
+semantic_search:
+  enabled: True
+  model_size: small
+```
 
-1. Semantic search is used in conjunction with the other filters available on the Search page. Use a combination of traditional filtering and semantic search for the best results.
-2. The comparison between text and image embedding distances generally means that results matching `description` will appear first, even if a `thumbnail` embedding may be a better match. Play with the "Search Type" filter to help find what you are looking for.
-3. Make your search language and tone closely match your descriptions. If you are using thumbnail search, phrase your query as an image caption.
-4. Semantic search on thumbnails tends to return better results when matching large subjects that take up most of the frame. Small things like "cat" tend to not work well.
-5. Experiment! Find a tracked object you want to test and start typing keywords to see what works for you.
+- Configuring the `large` model employs the full Jina model and will automatically run on the GPU if applicable.
+- Configuring the `small` model employs a quantized version of the model that uses less RAM and runs on CPU with a very negligible difference in embedding quality.
+
+### GPU Acceleration
+
+The CLIP models are downloaded in ONNX format, and the `large` model can be accelerated using GPU hardware, when available. This depends on the Docker build that is used.
+
+```yaml
+semantic_search:
+  enabled: True
+  model_size: large
+```
+
+:::info
+
+If the correct build is used for your GPU and the `large` model is configured, then the GPU will be detected and used automatically.
+
+**NOTE:** Object detection and Semantic Search are independent features. If you want to use your GPU with Semantic Search, you must choose the appropriate Frigate Docker image for your GPU.
+
+- **AMD**
+
+  - ROCm will automatically be detected and used for Semantic Search in the `-rocm` Frigate image.
+
+- **Intel**
+
+  - OpenVINO will automatically be detected and used for Semantic Search in the default Frigate image.
+
+- **Nvidia**
+  - Nvidia GPUs will automatically be detected and used for Semantic Search in the `-tensorrt` Frigate image.
+  - Jetson devices will automatically be detected and used for Semantic Search in the `-tensorrt-jp(4/5)` Frigate image.
+
+:::
+
+## Usage and Best Practices
+
+1. Semantic Search is used in conjunction with the other filters available on the Search page. Use a combination of traditional filtering and Semantic Search for the best results.
+2. Use the thumbnail search type when searching for particular objects in the scene. Use the description search type when attempting to discern the intent of your object.
+3. Because of how the AI models Frigate uses have been trained, the comparison between text and image embedding distances generally means that with multi-modal (`thumbnail` and `description`) searches, results matching `description` will appear first, even if a `thumbnail` embedding may be a better match. Play with the "Search Type" setting to help find what you are looking for. Note that if you are generating descriptions for specific objects or zones only, this may cause search results to prioritize the objects with descriptions even if the the ones without them are more relevant.
+4. Make your search language and tone closely match exactly what you're looking for. If you are using thumbnail search, **phrase your query as an image caption**. Searching for "red car" may not work as well as "red sedan driving down a residential street on a sunny day".
+5. Semantic search on thumbnails tends to return better results when matching large subjects that take up most of the frame. Small things like "cat" tend to not work well.
+6. Experiment! Find a tracked object you want to test and start typing keywords and phrases to see what works for you.

docs/docs/frigate/installation.md

@@ -81,15 +81,15 @@ You can calculate the **minimum** shm size for each camera with the following fo
 
 ```console
 # Replace <width> and <height>
-$ python -c 'print("{:.2f}MB".format((<width> * <height> * 1.5 * 10 + 270480) / 1048576))'
+$ python -c 'print("{:.2f}MB".format((<width> * <height> * 1.5 * 20 + 270480) / 1048576))'
 
-# Example for 1280x720
-$ python -c 'print("{:.2f}MB".format((1280 * 720 * 1.5 * 10 + 270480) / 1048576))'
-13.44MB
+# Example for 1280x720, including logs
+$ python -c 'print("{:.2f}MB".format((1280 * 720 * 1.5 * 20 + 270480) / 1048576)) + 40'
+46.63MB
 
 # Example for eight cameras detecting at 1280x720, including logs
-$ python -c 'print("{:.2f}MB".format(((1280 * 720 * 1.5 * 10 + 270480) / 1048576) * 8 + 40))'
-136.99MB
+$ python -c 'print("{:.2f}MB".format(((1280 * 720 * 1.5 * 20 + 270480) / 1048576) * 8 + 40))'
+253MB
 ```
 
 The shm size cannot be set per container for Home Assistant add-ons. However, this is probably not required since by default Home Assistant Supervisor allocates `/dev/shm` with half the size of your total memory. If your machine has 8GB of memory, chances are that Frigate will have access to up to 4GB without any additional configuration.
@@ -194,7 +194,7 @@ services:
     privileged: true # this may not be necessary for all setups
     restart: unless-stopped
     image: ghcr.io/blakeblackshear/frigate:stable
-    shm_size: "64mb" # update for your cameras based on calculation above
+    shm_size: "512mb" # update for your cameras based on calculation above
     devices:
       - /dev/bus/usb:/dev/bus/usb # Passes the USB Coral, needs to be modified for other versions
       - /dev/apex_0:/dev/apex_0 # Passes a PCIe Coral, follow driver instructions here https://coral.ai/docs/m2/get-started/#2a-on-linux
@@ -250,10 +250,7 @@ The community supported docker image tags for the current stable version are:
 - `stable-tensorrt-jp5` - Frigate build optimized for nvidia Jetson devices running Jetpack 5
 - `stable-tensorrt-jp4` - Frigate build optimized for nvidia Jetson devices running Jetpack 4.6
 - `stable-rk` - Frigate build for SBCs with Rockchip SoC
-- `stable-rocm` - Frigate build for [AMD GPUs and iGPUs](../configuration/object_detectors.md#amdrocm-gpu-detector), all drivers
-  - `stable-rocm-gfx900` - AMD gfx900 driver only
-  - `stable-rocm-gfx1030` - AMD gfx1030 driver only
-  - `stable-rocm-gfx1100` - AMD gfx1100 driver only
+- `stable-rocm` - Frigate build for [AMD GPUs](../configuration/object_detectors.md#amdrocm-gpu-detector)
   - `stable-h8l` - Frigate build for the Hailo-8L M.2 PICe Raspberry Pi 5 hat
 
 ## Home Assistant Addon

docs/docs/guides/configuring_go2rtc.md

@@ -13,7 +13,15 @@ Use of the bundled go2rtc is optional. You can still configure FFmpeg to connect
 
 # Setup a go2rtc stream
 
-First, you will want to configure go2rtc to connect to your camera stream by adding the stream you want to use for live view in your Frigate config file. For the best experience, you should set the stream name under go2rtc to match the name of your camera so that Frigate will automatically map it and be able to use better live view options for the camera. Avoid changing any other parts of your config at this step. Note that go2rtc supports [many different stream types](https://github.com/AlexxIT/go2rtc/tree/v1.9.4#module-streams), not just rtsp.
+First, you will want to configure go2rtc to connect to your camera stream by adding the stream you want to use for live view in your Frigate config file. Avoid changing any other parts of your config at this step. Note that go2rtc supports [many different stream types](https://github.com/AlexxIT/go2rtc/tree/v1.9.2#module-streams), not just rtsp.
+
+:::tip
+
+For the best experience, you should set the stream name under `go2rtc` to match the name of your camera so that Frigate will automatically map it and be able to use better live view options for the camera.
+
+See [the live view docs](../configuration/live.md#setting-stream-for-live-ui) for more information.
+
+:::
 
 ```yaml
 go2rtc:
@@ -39,8 +47,8 @@ After adding this to the config, restart Frigate and try to watch the live strea
 
 - Check Video Codec:
     - If the camera stream works in go2rtc but not in your browser, the video codec might be unsupported.
-    - If using H265, switch to H264. Refer to [video codec compatibility](https://github.com/AlexxIT/go2rtc/tree/v1.9.4#codecs-madness) in go2rtc documentation.
-    - If unable to switch from H265 to H264, or if the stream format is different (e.g., MJPEG), re-encode the video using [FFmpeg parameters](https://github.com/AlexxIT/go2rtc/tree/v1.9.4#source-ffmpeg). It supports rotating and resizing video feeds and hardware acceleration. Keep in mind that transcoding video from one format to another is a resource intensive task and you may be better off using the built-in jsmpeg view.
+    - If using H265, switch to H264. Refer to [video codec compatibility](https://github.com/AlexxIT/go2rtc/tree/v1.9.2#codecs-madness) in go2rtc documentation.
+    - If unable to switch from H265 to H264, or if the stream format is different (e.g., MJPEG), re-encode the video using [FFmpeg parameters](https://github.com/AlexxIT/go2rtc/tree/v1.9.2#source-ffmpeg). It supports rotating and resizing video feeds and hardware acceleration. Keep in mind that transcoding video from one format to another is a resource intensive task and you may be better off using the built-in jsmpeg view.
         ```yaml
         go2rtc:
           streams:

docs/docs/guides/getting_started.md

@@ -306,7 +306,9 @@ By default, Frigate will retain video of all tracked objects for 10 days. The fu
 
 ### Step 7: Complete config
 
-At this point you have a complete config with basic functionality. You can see the [full config reference](../configuration/reference.md) for a complete list of configuration options.
+At this point you have a complete config with basic functionality. 
+- View [common configuration examples](../configuration/index.md#common-configuration-examples) for a list of common configuration examples.
+- View [full config reference](../configuration/reference.md) for a complete list of configuration options.
 
 ### Follow up
 

docs/docs/integrations/mqtt.md

@@ -94,6 +94,18 @@ Message published for each changed tracked object. The first message is publishe
 }
 ```
 
+### `frigate/tracked_object_update`
+
+Message published for updates to tracked object metadata, for example when GenAI runs and returns a tracked object description.
+
+```json
+{
+  "type": "description",
+  "id": "1607123955.475377-mxklsc",
+  "description": "The car is a red sedan moving away from the camera."
+}
+```
+
 ### `frigate/reviews`
 
 Message published for each changed review item. The first message is published when the `detection` or `alert` is initiated. When additional objects are detected or when a zone change occurs, it will publish a, `update` message with the same id. When the review activity has ended a final `end` message is published.

docs/docs/plus/first_model.md

@@ -5,7 +5,7 @@ title: Requesting your first model
 
 ## Step 1: Upload and annotate your images
 
-Before requesting your first model, you will need to upload at least 10 images to Frigate+. But for the best results, you should provide at least 100 verified images per camera. Keep in mind that varying conditions should be included. You will want images from cloudy days, sunny days, dawn, dusk, and night. Refer to the [integration docs](../integrations/plus.md#generate-an-api-key) for instructions on how to easily submit images to Frigate+ directly from Frigate.
+Before requesting your first model, you will need to upload and verify at least 1 image to Frigate+. The more images you upload, annotate, and verify the better your results will be. Most users start to see very good results once they have at least 100 verified images per camera. Keep in mind that varying conditions should be included. You will want images from cloudy days, sunny days, dawn, dusk, and night. Refer to the [integration docs](../integrations/plus.md#generate-an-api-key) for instructions on how to easily submit images to Frigate+ directly from Frigate.
 
 It is recommended to submit **both** true positives and false positives. This will help the model differentiate between what is and isn't correct. You should aim for a target of 80% true positive submissions and 20% false positives across all of your images. If you are experiencing false positives in a specific area, submitting true positives for any object type near that area in similar lighting conditions will help teach the model what that area looks like when no objects are present.
 
@@ -13,7 +13,7 @@ For more detailed recommendations, you can refer to the docs on [improving your
 
 ## Step 2: Submit a model request
 
-Once you have an initial set of verified images, you can request a model on the Models page. Each model request requires 1 of the 12 trainings that you receive with your annual subscription. This model will support all [label types available](./index.md#available-label-types) even if you do not submit any examples for those labels. Model creation can take up to 36 hours.
+Once you have an initial set of verified images, you can request a model on the Models page. For guidance on choosing a model type, refer to [this part of the documentation](./index.md#available-model-types). Each model request requires 1 of the 12 trainings that you receive with your annual subscription. This model will support all [label types available](./index.md#available-label-types) even if you do not submit any examples for those labels. Model creation can take up to 36 hours.
 ![Plus Models Page](/img/plus/plus-models.jpg)
 
 ## Step 3: Set your model id in the config

docs/docs/plus/improving_model.md

@@ -3,7 +3,7 @@ id: improving_model
 title: Improving your model
 ---
 
-You may find that Frigate+ models result in more false positives initially, but by submitting true and false positives, the model will improve. Because a limited number of users submitted images to Frigate+ prior to this launch, you may need to submit several hundred images per camera to see good results. With all the new images now being submitted, future base models will improve as more and more users (including you) submit examples to Frigate+. Note that only verified images will be used when training your model. Submitting an image from Frigate as a true or false positive will not verify the image. You still must verify the image in Frigate+ in order for it to be used in training.
+You may find that Frigate+ models result in more false positives initially, but by submitting true and false positives, the model will improve. With all the new images now being submitted by subscribers, future base models will improve as more and more examples are incorporated. Note that only images with at least one verified label will be used when training your model. Submitting an image from Frigate as a true or false positive will not verify the image. You still must verify the image in Frigate+ in order for it to be used in training.
 
 - **Submit both true positives and false positives**. This will help the model differentiate between what is and isn't correct. You should aim for a target of 80% true positive submissions and 20% false positives across all of your images. If you are experiencing false positives in a specific area, submitting true positives for any object type near that area in similar lighting conditions will help teach the model what that area looks like when no objects are present.
 - **Lower your thresholds a little in order to generate more false/true positives near the threshold value**. For example, if you have some false positives that are scoring at 68% and some true positives scoring at 72%, you can try lowering your threshold to 65% and submitting both true and false positives within that range. This will help the model learn and widen the gap between true and false positive scores.
@@ -36,18 +36,17 @@ Misidentified objects should have a correct label added. For example, if a perso
 
 ## Shortcuts for a faster workflow
 
-|Shortcut Key|Description|
-|-----|--------|
-|`?`|Show all keyboard shortcuts|
-|`w`|Add box|
-|`d`|Toggle difficult|
-|`s`|Switch to the next label|
-|`tab`|Select next largest box|
-|`del`|Delete current box|
-|`esc`|Deselect/Cancel|
-|`← ↑ → ↓`|Move box|
-|`Shift + ← ↑ → ↓`|Resize box|
-|`-`|Zoom out|
-|`=`|Zoom in|
-|`f`|Hide/show all but current box|
-|`spacebar`|Verify and save|
+| Shortcut Key      | Description                   |
+| ----------------- | ----------------------------- |
+| `?`               | Show all keyboard shortcuts   |
+| `w`               | Add box                       |
+| `d`               | Toggle difficult              |
+| `s`               | Switch to the next label      |
+| `tab`             | Select next largest box       |
+| `del`             | Delete current box            |
+| `esc`             | Deselect/Cancel               |
+| `← ↑ → ↓`         | Move box                      |
+| `Shift + ← ↑ → ↓` | Resize box                    |
+| `scrollwheel`     | Zoom in/out                   |
+| `f`               | Hide/show all but current box |
+| `spacebar`        | Verify and save               |

docs/docs/plus/index.md

@@ -15,17 +15,36 @@ With a subscription, 12 model trainings per year are included. If you cancel you
 
 Information on how to integrate Frigate+ with Frigate can be found in the [integration docs](../integrations/plus.md).
 
+## Available model types
+
+There are two model types offered in Frigate+: `mobiledet` and `yolonas`. Both of these models are object detection models and are trained to detect the same set of labels [listed below](#available-label-types).
+
+Not all model types are supported by all detectors, so it's important to choose a model type to match your detector as shown in the table under [supported detector types](#supported-detector-types).
+
+| Model Type  | Description                                                                                                                                  |
+| ----------- | -------------------------------------------------------------------------------------------------------------------------------------------- |
+| `mobiledet` | Based on the same architecture as the default model included with Frigate. Runs on Google Coral devices and CPUs.                            |
+| `yolonas`   | A newer architecture that offers slightly higher accuracy and improved detection of small objects. Runs on Intel, NVidia GPUs, and AMD GPUs. |
+
 ## Supported detector types
 
+Currently, Frigate+ models support CPU (`cpu`), Google Coral (`edgetpu`), OpenVino (`openvino`), ONNX (`onnx`), and ROCm (`rocm`) detectors.
+
 :::warning
 
-Frigate+ models are not supported for TensorRT or OpenVino yet.
+Using Frigate+ models with `onnx` and `rocm` is only available with Frigate 0.15, which is still under development.
 
 :::
 
-Currently, Frigate+ models only support CPU (`cpu`) and Coral (`edgetpu`) models. OpenVino is next in line to gain support.
+| Hardware                                                                                                                     | Recommended Detector Type | Recommended Model Type |
+| ---------------------------------------------------------------------------------------------------------------------------- | ------------------------- | ---------------------- |
+| [CPU](/configuration/object_detectors.md#cpu-detector-not-recommended)                                                       | `cpu`                     | `mobiledet`            |
+| [Coral (all form factors)](/configuration/object_detectors.md#edge-tpu-detector)                                             | `edgetpu`                 | `mobiledet`            |
+| [Intel](/configuration/object_detectors.md#openvino-detector)                                                                | `openvino`                | `yolonas`              |
+| [NVidia GPU](https://deploy-preview-13787--frigate-docs.netlify.app/configuration/object_detectors#onnx)\*                   | `onnx`                    | `yolonas`              |
+| [AMD ROCm GPU](https://deploy-preview-13787--frigate-docs.netlify.app/configuration/object_detectors#amdrocm-gpu-detector)\* | `rocm`                    | `yolonas`              |
 
-The models are created using the same MobileDet architecture as the default model. Additional architectures will be added in future releases as needed.
+_\* Requires Frigate 0.15_
 
 ## Available label types
 

docs/docs/troubleshooting/edgetpu.md

@@ -49,7 +49,10 @@ The USB Coral can become stuck and need to be restarted, this can happen for a n
 
 ## PCIe Coral Not Detected
 
-The most common reason for the PCIe coral not being detected is that the driver has not been installed. See [the coral docs](https://coral.ai/docs/m2/get-started/#2-install-the-pcie-driver-and-edge-tpu-runtime) for how to install the driver for the PCIe based coral.
+The most common reason for the PCIe Coral not being detected is that the driver has not been installed. This process varies based on what OS and kernel that is being run. 
+
+- In most cases [the Coral docs](https://coral.ai/docs/m2/get-started/#2-install-the-pcie-driver-and-edge-tpu-runtime) show how to install the driver for the PCIe based Coral.
+- For Ubuntu 22.04+ https://github.com/jnicolson/gasket-builder can be used to build and install the latest version of the driver.
 
 ## Only One PCIe Coral Is Detected With Coral Dual EdgeTPU
 

docs/sidebars.ts

@@ -26,7 +26,7 @@ const sidebars: SidebarsConfig = {
         {
           type: 'link',
           label: 'Go2RTC Configuration Reference',
-          href: 'https://github.com/AlexxIT/go2rtc/tree/v1.9.4#configuration',
+          href: 'https://github.com/AlexxIT/go2rtc/tree/v1.9.2#configuration',
         } as PropSidebarItemLink,
       ],
       Detectors: [

docs/static/frigate-api.yaml

@@ -7,7 +7,7 @@ info:
 
 servers:
   - url: https://demo.frigate.video/api
-  - url: http://localhost:5001/
+  - url: http://localhost:5001/api
 
 paths:
   /auth:
@@ -172,76 +172,65 @@ paths:
           in: query
           required: false
           schema:
-            anyOf:
-              - type: string
-              - type: 'null'
+            type: string
             default: all
             title: Cameras
         - name: labels
           in: query
           required: false
           schema:
-            anyOf:
-              - type: string
-              - type: 'null'
+            type: string
             default: all
             title: Labels
         - name: zones
           in: query
           required: false
           schema:
-            anyOf:
-              - type: string
-              - type: 'null'
+            type: string
             default: all
             title: Zones
         - name: reviewed
           in: query
           required: false
           schema:
-            anyOf:
-              - type: integer
-              - type: 'null'
+            type: integer
             default: 0
             title: Reviewed
         - name: limit
           in: query
           required: false
           schema:
-            anyOf:
-              - type: integer
-              - type: 'null'
+            type: integer
             title: Limit
         - name: severity
           in: query
           required: false
           schema:
-            anyOf:
-              - type: string
-              - type: 'null'
+            allOf:
+              - $ref: '#/components/schemas/SeverityEnum'
             title: Severity
         - name: before
           in: query
           required: false
           schema:
-            anyOf:
-              - type: number
-              - type: 'null'
+            type: number
             title: Before
         - name: after
           in: query
           required: false
           schema:
-            anyOf:
-              - type: number
-              - type: 'null'
+            type: number
             title: After
       responses:
         '200':
           description: Successful Response
           content:
             application/json:
-              schema: { }
+              schema:
+                type: array
+                items:
+                  $ref: '#/components/schemas/ReviewSegmentResponse'
+                title: Response Review Review Get
         '422':
           description: Validation Error
           content:
@@ -259,36 +248,28 @@ paths:
           in: query
           required: false
           schema:
-            anyOf:
-              - type: string
-              - type: 'null'
+            type: string
             default: all
             title: Cameras
         - name: labels
           in: query
           required: false
           schema:
-            anyOf:
-              - type: string
-              - type: 'null'
+            type: string
             default: all
             title: Labels
         - name: zones
           in: query
           required: false
           schema:
-            anyOf:
-              - type: string
-              - type: 'null'
+            type: string
             default: all
             title: Zones
         - name: timezone
           in: query
           required: false
           schema:
-            anyOf:
-              - type: string
-              - type: 'null'
+            type: string
             default: utc
             title: Timezone
       responses:
@@ -296,7 +277,8 @@ paths:
           description: Successful Response
           content:
             application/json:
-              schema: { }
+              schema:
+                $ref: '#/components/schemas/ReviewSummaryResponse'
         '422':
           description: Validation Error
           content:
@@ -310,17 +292,18 @@ paths:
       summary: Set Multiple Reviewed
       operationId: set_multiple_reviewed_reviews_viewed_post
       requestBody:
+        required: true
         content:
           application/json:
             schema:
-              type: object
-              title: Body
+              $ref: '#/components/schemas/ReviewModifyMultipleBody'
       responses:
         '200':
           description: Successful Response
           content:
             application/json:
-              schema: { }
+              schema:
+                $ref: '#/components/schemas/GenericResponse'
         '422':
           description: Validation Error
           content:
@@ -334,17 +317,18 @@ paths:
       summary: Delete Reviews
       operationId: delete_reviews_reviews_delete_post
       requestBody:
+        required: true
         content:
           application/json:
             schema:
-              type: object
-              title: Body
+              $ref: '#/components/schemas/ReviewModifyMultipleBody'
       responses:
         '200':
           description: Successful Response
           content:
             application/json:
-              schema: { }
+              schema:
+                $ref: '#/components/schemas/GenericResponse'
         '422':
           description: Validation Error
           content:
@@ -363,96 +347,38 @@ paths:
           in: query
           required: false
           schema:
-            anyOf:
-              - type: string
-              - type: 'null'
+            type: string
             default: all
             title: Cameras
         - name: before
           in: query
           required: false
           schema:
-            anyOf:
-              - type: number
-              - type: 'null'
+            type: number
             title: Before
         - name: after
           in: query
           required: false
           schema:
-            anyOf:
-              - type: number
-              - type: 'null'
+            type: number
             title: After
         - name: scale
           in: query
           required: false
           schema:
-            anyOf:
-              - type: integer
-              - type: 'null'
+            type: integer
             default: 30
             title: Scale
       responses:
         '200':
           description: Successful Response
-          content:
-            application/json:
-              schema: { }
-        '422':
-          description: Validation Error
           content:
             application/json:
               schema:
-                $ref: '#/components/schemas/HTTPValidationError'
-  /review/activity/audio:
-    get:
-      tags:
-        - Review
-      summary: Audio Activity
-      description: Get motion and audio activity.
-      operationId: audio_activity_review_activity_audio_get
-      parameters:
-        - name: cameras
-          in: query
-          required: false
-          schema:
-            anyOf:
-              - type: string
-              - type: 'null'
-            default: all
-            title: Cameras
-        - name: before
-          in: query
-          required: false
-          schema:
-            anyOf:
-              - type: number
-              - type: 'null'
-            title: Before
-        - name: after
-          in: query
-          required: false
-          schema:
-            anyOf:
-              - type: number
-              - type: 'null'
-            title: After
-        - name: scale
-          in: query
-          required: false
-          schema:
-            anyOf:
-              - type: integer
-              - type: 'null'
-            default: 30
-            title: Scale
-      responses:
-        '200':
-          description: Successful Response
-          content:
-            application/json:
-              schema: { }
+                type: array
+                items:
+                  $ref: '#/components/schemas/ReviewActivityMotionResponse'
+                title: Response Motion Activity Review Activity Motion Get
         '422':
           description: Validation Error
           content:
@@ -477,57 +403,60 @@ paths:
           description: Successful Response
           content:
             application/json:
-              schema: { }
+              schema:
+                $ref: '#/components/schemas/ReviewSegmentResponse'
         '422':
           description: Validation Error
           content:
             application/json:
               schema:
                 $ref: '#/components/schemas/HTTPValidationError'
-  /review/{event_id}:
+  /review/{review_id}:
     get:
       tags:
         - Review
       summary: Get Review
-      operationId: get_review_review__event_id__get
+      operationId: get_review_review__review_id__get
       parameters:
-        - name: event_id
+        - name: review_id
           in: path
           required: true
           schema:
             type: string
-            title: Event Id
+            title: Review Id
       responses:
         '200':
           description: Successful Response
           content:
             application/json:
-              schema: { }
+              schema:
+                $ref: '#/components/schemas/ReviewSegmentResponse'
         '422':
           description: Validation Error
           content:
             application/json:
               schema:
                 $ref: '#/components/schemas/HTTPValidationError'
-  /review/{event_id}/viewed:
+  /review/{review_id}/viewed:
     delete:
       tags:
         - Review
       summary: Set Not Reviewed
-      operationId: set_not_reviewed_review__event_id__viewed_delete
+      operationId: set_not_reviewed_review__review_id__viewed_delete
       parameters:
-        - name: event_id
+        - name: review_id
           in: path
           required: true
           schema:
             type: string
-            title: Event Id
+            title: Review Id
       responses:
         '200':
           description: Successful Response
           content:
             application/json:
-              schema: { }
+              schema:
+                $ref: '#/components/schemas/GenericResponse'
         '422':
           description: Validation Error
           content:
@@ -763,13 +692,25 @@ paths:
           content:
             application/json:
               schema: { }
+  /nvinfo:
+    get:
+      tags:
+        - App
+      summary: Nvinfo
+      operationId: nvinfo_nvinfo_get
+      responses:
+        '200':
+          description: Successful Response
+          content:
+            application/json:
+              schema: { }
   /logs/{service}:
     get:
       tags:
         - App
         - Logs
       summary: Logs
-      description: Get logs for the requested service (frigate/nginx/go2rtc/chroma)
+      description: Get logs for the requested service (frigate/nginx/go2rtc)
       operationId: logs_logs__service__get
       parameters:
         - name: service
@@ -781,7 +722,6 @@ paths:
               - frigate
               - nginx
               - go2rtc
-              - chroma
             title: Service
         - name: download
           in: query
@@ -1042,7 +982,8 @@ paths:
         - Preview
       summary: Preview Hour
       description: Get all mp4 previews relevant for time period given the timezone
-      operationId: preview_hour_preview__year_month___day___hour___camera_name___tz_name__get
+      operationId: >-
+        preview_hour_preview__year_month___day___hour___camera_name___tz_name__get
       parameters:
         - name: year_month
           in: path
@@ -1092,7 +1033,8 @@ paths:
         - Preview
       summary: Get Preview Frames From Cache
       description: Get list of cached preview frames
-      operationId: get_preview_frames_from_cache_preview__camera_name__start__start_ts__end__end_ts__frames_get
+      operationId: >-
+        get_preview_frames_from_cache_preview__camera_name__start__start_ts__end__end_ts__frames_get
       parameters:
         - name: camera_name
           in: path
@@ -1177,7 +1119,8 @@ paths:
       tags:
         - Export
       summary: Export Recording
-      operationId: export_recording_export__camera_name__start__start_time__end__end_time__post
+      operationId: >-
+        export_recording_export__camera_name__start__start_time__end__end_time__post
       parameters:
         - name: camera_name
           in: path
@@ -1198,11 +1141,11 @@ paths:
             type: number
             title: End Time
       requestBody:
+        required: true
         content:
           application/json:
             schema:
-              type: object
-              title: Body
+              $ref: '#/components/schemas/ExportRecordingsBody'
       responses:
         '200':
           description: Successful Response
@@ -1465,6 +1408,14 @@ paths:
               - type: number
               - type: 'null'
             title: Max Length
+        - name: event_id
+          in: query
+          required: false
+          schema:
+            anyOf:
+              - type: string
+              - type: 'null'
+            title: Event Id
         - name: sort
           in: query
           required: false
@@ -1575,7 +1526,7 @@ paths:
             anyOf:
               - type: string
               - type: 'null'
-            default: thumbnail,description
+            default: thumbnail
             title: Search Type
         - name: include_thumbnails
           in: query
@@ -1647,6 +1598,22 @@ paths:
               - type: 'null'
             default: 00:00,24:00
             title: Time Range
+        - name: has_clip
+          in: query
+          required: false
+          schema:
+            anyOf:
+              - type: boolean
+              - type: 'null'
+            title: Has Clip
+        - name: has_snapshot
+          in: query
+          required: false
+          schema:
+            anyOf:
+              - type: boolean
+              - type: 'null'
+            title: Has Snapshot
         - name: timezone
           in: query
           required: false
@@ -1656,6 +1623,30 @@ paths:
               - type: 'null'
             default: utc
             title: Timezone
+        - name: min_score
+          in: query
+          required: false
+          schema:
+            anyOf:
+              - type: number
+              - type: 'null'
+            title: Min Score
+        - name: max_score
+          in: query
+          required: false
+          schema:
+            anyOf:
+              - type: number
+              - type: 'null'
+            title: Max Score
+        - name: sort
+          in: query
+          required: false
+          schema:
+            anyOf:
+              - type: string
+              - type: 'null'
+            title: Sort
       responses:
         '200':
           description: Successful Response
@@ -1942,6 +1933,15 @@ paths:
           schema:
             type: string
             title: Event Id
+        - name: source
+          in: query
+          required: false
+          schema:
+            anyOf:
+              - $ref: '#/components/schemas/RegenerateDescriptionEnum'
+              - type: 'null'
+            default: thumbnails
+            title: Source
       responses:
         '200':
           description: Successful Response
@@ -2029,12 +2029,12 @@ paths:
             application/json:
               schema:
                 $ref: '#/components/schemas/HTTPValidationError'
-  '{camera_name}':
+  /{camera_name}:
     get:
       tags:
         - Media
       summary: Mjpeg Feed
-      operationId: mjpeg_feed_camera_name__get
+      operationId: mjpeg_feed__camera_name__get
       parameters:
         - name: camera_name
           in: path
@@ -2241,7 +2241,8 @@ paths:
       tags:
         - Media
       summary: Get Snapshot From Recording
-      operationId: get_snapshot_from_recording__camera_name__recordings__frame_time__snapshot__format__get
+      operationId: >-
+        get_snapshot_from_recording__camera_name__recordings__frame_time__snapshot__format__get
       parameters:
         - name: camera_name
           in: path
@@ -2363,7 +2364,9 @@ paths:
       tags:
         - Media
       summary: Recordings
-      description: Return specific camera recordings between the given 'after'/'end' times. If not provided the last hour will be used
+      description: >-
+        Return specific camera recordings between the given 'after'/'end' times.
+        If not provided the last hour will be used
       operationId: recordings__camera_name__recordings_get
       parameters:
         - name: camera_name
@@ -2377,14 +2380,14 @@ paths:
           required: false
           schema:
             type: number
-            default: 1727542549.303557
+            default: 1731275308.238304
             title: After
         - name: before
           in: query
           required: false
           schema:
             type: number
-            default: 1727546149.303926
+            default: 1731278908.238313
             title: Before
       responses:
         '200':
@@ -2423,13 +2426,6 @@ paths:
           schema:
             type: number
             title: End Ts
-        - name: download
-          in: query
-          required: false
-          schema:
-            type: boolean
-            default: false
-            title: Download
       responses:
         '200':
           description: Successful Response
@@ -2800,13 +2796,6 @@ paths:
           schema:
             type: string
             title: Event Id
-        - name: download
-          in: query
-          required: false
-          schema:
-            type: boolean
-            default: false
-            title: Download
       responses:
         '200':
           description: Successful Response
@@ -3121,7 +3110,9 @@ paths:
       tags:
         - Media
       summary: Label Snapshot
-      description: Returns the snapshot image from the latest event for the given camera and label combo
+      description: >-
+        Returns the snapshot image from the latest event for the given camera
+        and label combo
       operationId: label_snapshot__camera_name___label__snapshot_jpg_get
       parameters:
         - name: camera_name
@@ -3193,6 +3184,32 @@ components:
       required:
         - password
       title: AppPutPasswordBody
+    DayReview:
+      properties:
+        day:
+          type: string
+          format: date-time
+          title: Day
+        reviewed_alert:
+          type: integer
+          title: Reviewed Alert
+        reviewed_detection:
+          type: integer
+          title: Reviewed Detection
+        total_alert:
+          type: integer
+          title: Total Alert
+        total_detection:
+          type: integer
+          title: Total Detection
+      type: object
+      required:
+        - day
+        - reviewed_alert
+        - reviewed_detection
+        - total_alert
+        - total_detection
+      title: DayReview
     EventsCreateBody:
       properties:
         source_type:
@@ -3237,7 +3254,6 @@ components:
         description:
           anyOf:
             - type: string
-              minLength: 1
             - type: 'null'
           title: The description of the event
       type: object
@@ -3270,6 +3286,27 @@ components:
       required:
         - subLabel
       title: EventsSubLabelBody
+    ExportRecordingsBody:
+      properties:
+        playback:
+          allOf:
+            - $ref: '#/components/schemas/PlaybackFactorEnum'
+          title: Playback factor
+          default: realtime
+        source:
+          allOf:
+            - $ref: '#/components/schemas/PlaybackSourceEnum'
+          title: Playback source
+          default: recordings
+        name:
+          type: string
+          maxLength: 256
+          title: Friendly name
+        image_path:
+          type: string
+          title: Image Path
+      type: object
+      title: ExportRecordingsBody
     Extension:
       type: string
       enum:
@@ -3278,6 +3315,19 @@ components:
         - jpg
         - jpeg
       title: Extension
+    GenericResponse:
+      properties:
+        success:
+          type: boolean
+          title: Success
+        message:
+          type: string
+          title: Message
+      type: object
+      required:
+        - success
+        - message
+      title: GenericResponse
     HTTPValidationError:
       properties:
         detail:
@@ -3287,6 +3337,132 @@ components:
           title: Detail
       type: object
       title: HTTPValidationError
+    Last24HoursReview:
+      properties:
+        reviewed_alert:
+          type: integer
+          title: Reviewed Alert
+        reviewed_detection:
+          type: integer
+          title: Reviewed Detection
+        total_alert:
+          type: integer
+          title: Total Alert
+        total_detection:
+          type: integer
+          title: Total Detection
+      type: object
+      required:
+        - reviewed_alert
+        - reviewed_detection
+        - total_alert
+        - total_detection
+      title: Last24HoursReview
+    PlaybackFactorEnum:
+      type: string
+      enum:
+        - realtime
+        - timelapse_25x
+      title: PlaybackFactorEnum
+    PlaybackSourceEnum:
+      type: string
+      enum:
+        - recordings
+        - preview
+      title: PlaybackSourceEnum
+    RegenerateDescriptionEnum:
+      type: string
+      enum:
+        - thumbnails
+        - snapshot
+      title: RegenerateDescriptionEnum
+    ReviewActivityMotionResponse:
+      properties:
+        start_time:
+          type: integer
+          title: Start Time
+        motion:
+          type: number
+          title: Motion
+        camera:
+          type: string
+          title: Camera
+      type: object
+      required:
+        - start_time
+        - motion
+        - camera
+      title: ReviewActivityMotionResponse
+    ReviewModifyMultipleBody:
+      properties:
+        ids:
+          items:
+            type: string
+            minLength: 1
+          type: array
+          minItems: 1
+          title: Ids
+      type: object
+      required:
+        - ids
+      title: ReviewModifyMultipleBody
+    ReviewSegmentResponse:
+      properties:
+        id:
+          type: string
+          title: Id
+        camera:
+          type: string
+          title: Camera
+        start_time:
+          type: string
+          format: date-time
+          title: Start Time
+        end_time:
+          type: string
+          format: date-time
+          title: End Time
+        has_been_reviewed:
+          type: boolean
+          title: Has Been Reviewed
+        severity:
+          $ref: '#/components/schemas/SeverityEnum'
+        thumb_path:
+          type: string
+          title: Thumb Path
+        data:
+          title: Data
+      type: object
+      required:
+        - id
+        - camera
+        - start_time
+        - end_time
+        - has_been_reviewed
+        - severity
+        - thumb_path
+        - data
+      title: ReviewSegmentResponse
+    ReviewSummaryResponse:
+      properties:
+        last24Hours:
+          $ref: '#/components/schemas/Last24HoursReview'
+        root:
+          additionalProperties:
+            $ref: '#/components/schemas/DayReview'
+          type: object
+          title: Root
+      type: object
+      required:
+        - last24Hours
+        - root
+      title: ReviewSummaryResponse
+    SeverityEnum:
+      type: string
+      enum:
+        - alert
+        - detection
+      title: SeverityEnum
     SubmitPlusBody:
       properties:
         include_annotation:

frigate/api/auth.py

@@ -357,6 +357,7 @@ def create_user(request: Request, body: AppPostUsersBody):
         {
             User.username: body.username,
             User.password_hash: password_hash,
+            User.notification_tokens: [],
         }
     ).execute()
     return JSONResponse(content={"username": body.username})

frigate/api/defs/events_body.py

@@ -11,9 +11,7 @@ class EventsSubLabelBody(BaseModel):
 
 
 class EventsDescriptionBody(BaseModel):
-    description: Union[str, None] = Field(
-        title="The description of the event", min_length=1
-    )
+    description: Union[str, None] = Field(title="The description of the event")
 
 
 class EventsCreateBody(BaseModel):

frigate/api/defs/events_query_parameters.py

@@ -28,6 +28,7 @@ class EventsQueryParams(BaseModel):
     is_submitted: Optional[int] = None
     min_length: Optional[float] = None
     max_length: Optional[float] = None
+    event_id: Optional[str] = None
     sort: Optional[str] = None
     timezone: Optional[str] = "utc"
 
@@ -35,7 +36,7 @@ class EventsQueryParams(BaseModel):
 class EventsSearchQueryParams(BaseModel):
     query: Optional[str] = None
     event_id: Optional[str] = None
-    search_type: Optional[str] = "thumbnail,description"
+    search_type: Optional[str] = "thumbnail"
     include_thumbnails: Optional[int] = 1
     limit: Optional[int] = 50
     cameras: Optional[str] = "all"
@@ -44,7 +45,13 @@ class EventsSearchQueryParams(BaseModel):
     after: Optional[float] = None
     before: Optional[float] = None
     time_range: Optional[str] = DEFAULT_TIME_RANGE
+    has_clip: Optional[bool] = None
+    has_snapshot: Optional[bool] = None
+    is_submitted: Optional[bool] = None
     timezone: Optional[str] = "utc"
+    min_score: Optional[float] = None
+    max_score: Optional[float] = None
+    sort: Optional[str] = None
 
 
 class EventsSummaryQueryParams(BaseModel):

frigate/api/defs/generic_response.py

@@ -0,0 +1,6 @@
+from pydantic import BaseModel
+
+
+class GenericResponse(BaseModel):
+    success: bool
+    message: str

frigate/api/defs/request/export_recordings_body.py

@@ -0,0 +1,20 @@
+from typing import Union
+
+from pydantic import BaseModel, Field
+from pydantic.json_schema import SkipJsonSchema
+
+from frigate.record.export import (
+    PlaybackFactorEnum,
+    PlaybackSourceEnum,
+)
+
+
+class ExportRecordingsBody(BaseModel):
+    playback: PlaybackFactorEnum = Field(
+        default=PlaybackFactorEnum.realtime, title="Playback factor"
+    )
+    source: PlaybackSourceEnum = Field(
+        default=PlaybackSourceEnum.recordings, title="Playback source"
+    )
+    name: str = Field(title="Friendly name", default=None, max_length=256)
+    image_path: Union[str, SkipJsonSchema[None]] = None

frigate/api/defs/review_body.py

@@ -0,0 +1,6 @@
+from pydantic import BaseModel, conlist, constr
+
+
+class ReviewModifyMultipleBody(BaseModel):
+    # List of string with at least one element and each element with at least one char
+    ids: conlist(constr(min_length=1), min_length=1)

frigate/api/defs/review_query_parameters.py

@@ -1,28 +1,31 @@
-from typing import Optional
+from typing import Union
 
 from pydantic import BaseModel
+from pydantic.json_schema import SkipJsonSchema
+
+from frigate.review.maintainer import SeverityEnum
 
 
 class ReviewQueryParams(BaseModel):
-    cameras: Optional[str] = "all"
-    labels: Optional[str] = "all"
-    zones: Optional[str] = "all"
-    reviewed: Optional[int] = 0
-    limit: Optional[int] = None
-    severity: Optional[str] = None
-    before: Optional[float] = None
-    after: Optional[float] = None
+    cameras: str = "all"
+    labels: str = "all"
+    zones: str = "all"
+    reviewed: int = 0
+    limit: Union[int, SkipJsonSchema[None]] = None
+    severity: Union[SeverityEnum, SkipJsonSchema[None]] = None
+    before: Union[float, SkipJsonSchema[None]] = None
+    after: Union[float, SkipJsonSchema[None]] = None
 
 
 class ReviewSummaryQueryParams(BaseModel):
-    cameras: Optional[str] = "all"
-    labels: Optional[str] = "all"
-    zones: Optional[str] = "all"
-    timezone: Optional[str] = "utc"
+    cameras: str = "all"
+    labels: str = "all"
+    zones: str = "all"
+    timezone: str = "utc"
 
 
 class ReviewActivityMotionQueryParams(BaseModel):
-    cameras: Optional[str] = "all"
-    before: Optional[float] = None
-    after: Optional[float] = None
-    scale: Optional[int] = 30
+    cameras: str = "all"
+    before: Union[float, SkipJsonSchema[None]] = None
+    after: Union[float, SkipJsonSchema[None]] = None
+    scale: int = 30

frigate/api/defs/review_responses.py

@@ -0,0 +1,43 @@
+from datetime import datetime
+from typing import Dict
+
+from pydantic import BaseModel, Json
+
+from frigate.review.maintainer import SeverityEnum
+
+
+class ReviewSegmentResponse(BaseModel):
+    id: str
+    camera: str
+    start_time: datetime
+    end_time: datetime
+    has_been_reviewed: bool
+    severity: SeverityEnum
+    thumb_path: str
+    data: Json
+
+
+class Last24HoursReview(BaseModel):
+    reviewed_alert: int
+    reviewed_detection: int
+    total_alert: int
+    total_detection: int
+
+
+class DayReview(BaseModel):
+    day: datetime
+    reviewed_alert: int
+    reviewed_detection: int
+    total_alert: int
+    total_detection: int
+
+
+class ReviewSummaryResponse(BaseModel):
+    last24Hours: Last24HoursReview
+    root: Dict[str, DayReview]
+
+
+class ReviewActivityMotionResponse(BaseModel):
+    start_time: int
+    motion: float
+    camera: str

frigate/api/event.py

@@ -88,6 +88,7 @@ def events(params: EventsQueryParams = Depends()):
     is_submitted = params.is_submitted
     min_length = params.min_length
     max_length = params.max_length
+    event_id = params.event_id
 
     sort = params.sort
 
@@ -230,6 +231,9 @@ def events(params: EventsQueryParams = Depends()):
         elif is_submitted > 0:
             clauses.append((Event.plus_id != ""))
 
+    if event_id is not None:
+        clauses.append((Event.id == event_id))
+
     if len(clauses) == 0:
         clauses.append((True))
 
@@ -259,66 +263,61 @@ def events(params: EventsQueryParams = Depends()):
 
 @router.get("/events/explore")
 def events_explore(limit: int = 10):
-    subquery = Event.select(
-        Event.id,
-        Event.camera,
-        Event.label,
-        Event.zones,
-        Event.start_time,
-        Event.end_time,
-        Event.has_clip,
-        Event.has_snapshot,
-        Event.plus_id,
-        Event.retain_indefinitely,
-        Event.sub_label,
-        Event.top_score,
-        Event.false_positive,
-        Event.box,
-        Event.data,
-        fn.rank()
-        .over(partition_by=[Event.label], order_by=[Event.start_time.desc()])
-        .alias("rank"),
-        fn.COUNT(Event.id).over(partition_by=[Event.label]).alias("event_count"),
-    ).alias("subquery")
+    # get distinct labels for all events
+    distinct_labels = Event.select(Event.label).distinct().order_by(Event.label)
 
-    query = (
-        Event.select(
-            subquery.c.id,
-            subquery.c.camera,
-            subquery.c.label,
-            subquery.c.zones,
-            subquery.c.start_time,
-            subquery.c.end_time,
-            subquery.c.has_clip,
-            subquery.c.has_snapshot,
-            subquery.c.plus_id,
-            subquery.c.retain_indefinitely,
-            subquery.c.sub_label,
-            subquery.c.top_score,
-            subquery.c.false_positive,
-            subquery.c.box,
-            subquery.c.data,
-            subquery.c.event_count,
-        )
-        .from_(subquery)
-        .where(subquery.c.rank <= limit)
-        .order_by(subquery.c.event_count.desc(), subquery.c.start_time.desc())
-        .dicts()
-    )
+    label_counts = {}
 
-    events = list(query.iterator())
+    def event_generator():
+        for label_obj in distinct_labels.iterator():
+            label = label_obj.label
 
-    processed_events = [
-        {k: v for k, v in event.items() if k != "data"}
-        | {
-            "data": {
-                k: v
-                for k, v in event["data"].items()
-                if k in ["type", "score", "top_score", "description"]
+            # get most recent events for this label
+            label_events = (
+                Event.select()
+                .where(Event.label == label)
+                .order_by(Event.start_time.desc())
+                .limit(limit)
+                .iterator()
+            )
+
+            # count total events for this label
+            label_counts[label] = Event.select().where(Event.label == label).count()
+
+            yield from label_events
+
+    def process_events():
+        for event in event_generator():
+            processed_event = {
+                "id": event.id,
+                "camera": event.camera,
+                "label": event.label,
+                "zones": event.zones,
+                "start_time": event.start_time,
+                "end_time": event.end_time,
+                "has_clip": event.has_clip,
+                "has_snapshot": event.has_snapshot,
+                "plus_id": event.plus_id,
+                "retain_indefinitely": event.retain_indefinitely,
+                "sub_label": event.sub_label,
+                "top_score": event.top_score,
+                "false_positive": event.false_positive,
+                "box": event.box,
+                "data": {
+                    k: v
+                    for k, v in event.data.items()
+                    if k in ["type", "score", "top_score", "description"]
+                },
+                "event_count": label_counts[event.label],
             }
-        }
-        for event in events
-    ]
+            yield processed_event
+
+    # convert iterator to list and sort
+    processed_events = sorted(
+        process_events(),
+        key=lambda x: (x["event_count"], x["start_time"]),
+        reverse=True,
+    )
 
     return JSONResponse(content=processed_events)
 
@@ -348,6 +347,7 @@ def events_search(request: Request, params: EventsSearchQueryParams = Depends())
     search_type = params.search_type
     include_thumbnails = params.include_thumbnails
     limit = params.limit
+    sort = params.sort
 
     # Filters
     cameras = params.cameras
@@ -355,7 +355,12 @@ def events_search(request: Request, params: EventsSearchQueryParams = Depends())
     zones = params.zones
     after = params.after
     before = params.before
+    min_score = params.min_score
+    max_score = params.max_score
     time_range = params.time_range
+    has_clip = params.has_clip
+    has_snapshot = params.has_snapshot
+    is_submitted = params.is_submitted
 
     # for similarity search
     event_id = params.event_id
@@ -394,6 +399,7 @@ def events_search(request: Request, params: EventsSearchQueryParams = Depends())
         Event.end_time,
         Event.has_clip,
         Event.has_snapshot,
+        Event.top_score,
         Event.data,
         Event.plus_id,
         ReviewSegment.thumb_path,
@@ -430,6 +436,26 @@ def events_search(request: Request, params: EventsSearchQueryParams = Depends())
     if before:
         event_filters.append((Event.start_time < before))
 
+    if has_clip is not None:
+        event_filters.append((Event.has_clip == has_clip))
+
+    if has_snapshot is not None:
+        event_filters.append((Event.has_snapshot == has_snapshot))
+
+    if is_submitted is not None:
+        if is_submitted == 0:
+            event_filters.append((Event.plus_id.is_null()))
+        elif is_submitted > 0:
+            event_filters.append((Event.plus_id != ""))
+
+    if min_score is not None and max_score is not None:
+        event_filters.append((Event.data["score"].between(min_score, max_score)))
+    else:
+        if min_score is not None:
+            event_filters.append((Event.data["score"] >= min_score))
+        if max_score is not None:
+            event_filters.append((Event.data["score"] <= max_score))
+
     if time_range != DEFAULT_TIME_RANGE:
         tz_name = params.timezone
         hour_modifier, minute_modifier, _ = get_tz_modifiers(tz_name)
@@ -472,13 +498,8 @@ def events_search(request: Request, params: EventsSearchQueryParams = Depends())
                 status_code=404,
             )
 
-        thumb_result = context.embeddings.search_thumbnail(search_event)
-        thumb_ids = dict(
-            zip(
-                [result[0] for result in thumb_result],
-                context.thumb_stats.normalize([result[1] for result in thumb_result]),
-            )
-        )
+        thumb_result = context.search_thumbnail(search_event)
+        thumb_ids = {result[0]: result[1] for result in thumb_result}
         search_results = {
             event_id: {"distance": distance, "source": "thumbnail"}
             for event_id, distance in thumb_ids.items()
@@ -486,15 +507,18 @@ def events_search(request: Request, params: EventsSearchQueryParams = Depends())
     else:
         search_types = search_type.split(",")
 
+        # only save stats for multi-modal searches
+        save_stats = "thumbnail" in search_types and "description" in search_types
+
         if "thumbnail" in search_types:
-            thumb_result = context.embeddings.search_thumbnail(query)
+            thumb_result = context.search_thumbnail(query)
+
+            thumb_distances = context.thumb_stats.normalize(
+                [result[1] for result in thumb_result], save_stats
+            )
+
             thumb_ids = dict(
-                zip(
-                    [result[0] for result in thumb_result],
-                    context.thumb_stats.normalize(
-                        [result[1] for result in thumb_result]
-                    ),
-                )
+                zip([result[0] for result in thumb_result], thumb_distances)
             )
             search_results.update(
                 {
@@ -504,13 +528,14 @@ def events_search(request: Request, params: EventsSearchQueryParams = Depends())
             )
 
         if "description" in search_types:
-            desc_result = context.embeddings.search_description(query)
-            desc_ids = dict(
-                zip(
-                    [result[0] for result in desc_result],
-                    context.desc_stats.normalize([result[1] for result in desc_result]),
-                )
+            desc_result = context.search_description(query)
+
+            desc_distances = context.desc_stats.normalize(
+                [result[1] for result in desc_result], save_stats
             )
+
+            desc_ids = dict(zip([result[0] for result in desc_result], desc_distances))
+
             for event_id, distance in desc_ids.items():
                 if (
                     event_id not in search_results
@@ -555,11 +580,19 @@ def events_search(request: Request, params: EventsSearchQueryParams = Depends())
 
         processed_events.append(processed_event)
 
-    # Sort by search distance if search_results are available, otherwise by start_time
+    # Sort by search distance if search_results are available, otherwise by start_time as default
     if search_results:
         processed_events.sort(key=lambda x: x.get("search_distance", float("inf")))
     else:
-        processed_events.sort(key=lambda x: x["start_time"], reverse=True)
+        if sort == "score_asc":
+            processed_events.sort(key=lambda x: x["score"])
+        elif sort == "score_desc":
+            processed_events.sort(key=lambda x: x["score"], reverse=True)
+        elif sort == "date_asc":
+            processed_events.sort(key=lambda x: x["start_time"])
+        else:
+            # "date_desc" default
+            processed_events.sort(key=lambda x: x["start_time"], reverse=True)
 
     # Limit the number of events returned
     processed_events = processed_events[:limit]
@@ -927,27 +960,19 @@ def set_description(
 
     new_description = body.description
 
-    if new_description is None or len(new_description) == 0:
-        return JSONResponse(
-            content=(
-                {
-                    "success": False,
-                    "message": "description cannot be empty",
-                }
-            ),
-            status_code=400,
-        )
-
     event.data["description"] = new_description
     event.save()
 
     # If semantic search is enabled, update the index
     if request.app.frigate_config.semantic_search.enabled:
         context: EmbeddingsContext = request.app.embeddings
-        context.embeddings.upsert_description(
-            event_id=event_id,
-            description=new_description,
-        )
+        if len(new_description) > 0:
+            context.update_description(
+                event_id,
+                new_description,
+            )
+        else:
+            context.db.delete_embeddings_description(event_ids=[event_id])
 
     response_message = (
         f"Event {event_id} description is now blank"
@@ -978,9 +1003,11 @@ def regenerate_description(
             status_code=404,
         )
 
+    camera_config = request.app.frigate_config.cameras[event.camera]
+
     if (
         request.app.frigate_config.semantic_search.enabled
-        and request.app.frigate_config.genai.enabled
+        and camera_config.genai.enabled
     ):
         request.app.event_metadata_updater.publish((event.id, params.source))
 
@@ -1001,7 +1028,7 @@ def regenerate_description(
         content=(
             {
                 "success": False,
-                "message": "Semantic search and generative AI are not enabled",
+                "message": "Semantic Search and Generative AI must be enabled to regenerate a description",
             }
         ),
         status_code=400,
@@ -1024,17 +1051,14 @@ def delete_event(request: Request, event_id: str):
         media.unlink(missing_ok=True)
         media = Path(f"{os.path.join(CLIPS_DIR, media_name)}-clean.png")
         media.unlink(missing_ok=True)
-    if event.has_clip:
-        media = Path(f"{os.path.join(CLIPS_DIR, media_name)}.mp4")
-        media.unlink(missing_ok=True)
 
     event.delete_instance()
     Timeline.delete().where(Timeline.source_id == event_id).execute()
     # If semantic search is enabled, update the index
     if request.app.frigate_config.semantic_search.enabled:
         context: EmbeddingsContext = request.app.embeddings
-        context.embeddings.delete_thumbnail(id=[event_id])
-        context.embeddings.delete_description(id=[event_id])
+        context.db.delete_embeddings_thumbnail(event_ids=[event_id])
+        context.db.delete_embeddings_description(event_ids=[event_id])
     return JSONResponse(
         content=({"success": True, "message": "Event " + event_id + " deleted"}),
         status_code=200,

frigate/api/export.py

@@ -4,17 +4,22 @@ import logging
 import random
 import string
 from pathlib import Path
-from typing import Optional
 
 import psutil
 from fastapi import APIRouter, Request
 from fastapi.responses import JSONResponse
 from peewee import DoesNotExist
 
+from frigate.api.defs.request.export_recordings_body import ExportRecordingsBody
 from frigate.api.defs.tags import Tags
 from frigate.const import EXPORT_DIR
-from frigate.models import Export, Recordings
-from frigate.record.export import PlaybackFactorEnum, RecordingExporter
+from frigate.models import Export, Previews, Recordings
+from frigate.record.export import (
+    PlaybackFactorEnum,
+    PlaybackSourceEnum,
+    RecordingExporter,
+)
+from frigate.util.builtin import is_current_hour
 
 logger = logging.getLogger(__name__)
 
@@ -33,7 +38,7 @@ def export_recording(
     camera_name: str,
     start_time: float,
     end_time: float,
-    body: dict = None,
+    body: ExportRecordingsBody,
 ):
     if not camera_name or not request.app.frigate_config.cameras.get(camera_name):
         return JSONResponse(
@@ -43,36 +48,52 @@ def export_recording(
             status_code=404,
         )
 
-    json: dict[str, any] = body or {}
-    playback_factor = json.get("playback", "realtime")
-    friendly_name: Optional[str] = json.get("name")
+    playback_factor = body.playback
+    playback_source = body.source
+    friendly_name = body.name
+    existing_image = body.image_path
 
-    if len(friendly_name or "") > 256:
-        return JSONResponse(
-            content=({"success": False, "message": "File name is too long."}),
-            status_code=401,
+    if playback_source == "recordings":
+        recordings_count = (
+            Recordings.select()
+            .where(
+                Recordings.start_time.between(start_time, end_time)
+                | Recordings.end_time.between(start_time, end_time)
+                | (
+                    (start_time > Recordings.start_time)
+                    & (end_time < Recordings.end_time)
+                )
+            )
+            .where(Recordings.camera == camera_name)
+            .count()
         )
 
-    existing_image = json.get("image_path")
-
-    recordings_count = (
-        Recordings.select()
-        .where(
-            Recordings.start_time.between(start_time, end_time)
-            | Recordings.end_time.between(start_time, end_time)
-            | ((start_time > Recordings.start_time) & (end_time < Recordings.end_time))
+        if recordings_count <= 0:
+            return JSONResponse(
+                content=(
+                    {"success": False, "message": "No recordings found for time range"}
+                ),
+                status_code=400,
+            )
+    else:
+        previews_count = (
+            Previews.select()
+            .where(
+                Previews.start_time.between(start_time, end_time)
+                | Previews.end_time.between(start_time, end_time)
+                | ((start_time > Previews.start_time) & (end_time < Previews.end_time))
+            )
+            .where(Previews.camera == camera_name)
+            .count()
         )
-        .where(Recordings.camera == camera_name)
-        .count()
-    )
 
-    if recordings_count <= 0:
-        return JSONResponse(
-            content=(
-                {"success": False, "message": "No recordings found for time range"}
-            ),
-            status_code=400,
-        )
+        if not is_current_hour(start_time) and previews_count <= 0:
+            return JSONResponse(
+                content=(
+                    {"success": False, "message": "No previews found for time range"}
+                ),
+                status_code=400,
+            )
 
     export_id = f"{camera_name}_{''.join(random.choices(string.ascii_lowercase + string.digits, k=6))}"
     exporter = RecordingExporter(
@@ -88,6 +109,11 @@ def export_recording(
             if playback_factor in PlaybackFactorEnum.__members__.values()
             else PlaybackFactorEnum.realtime
         ),
+        (
+            PlaybackSourceEnum[playback_source]
+            if playback_source in PlaybackSourceEnum.__members__.values()
+            else PlaybackSourceEnum.recordings
+        ),
     )
     exporter.start()
     return JSONResponse(

frigate/api/fastapi_app.py

@@ -82,6 +82,10 @@ def create_fastapi_app(
             database.close()
         return response
 
+    @app.on_event("startup")
+    async def startup():
+        logger.info("FastAPI started")
+
     # Rate limiter (used for login endpoint)
     auth.rateLimiter.set_limit(frigate_config.auth.failed_login_rate_limit or "")
     app.state.limiter = limiter

frigate/api/media.py

@@ -7,6 +7,7 @@ import os
 import subprocess as sp
 import time
 from datetime import datetime, timedelta, timezone
+from pathlib import Path as FilePath
 from urllib.parse import unquote
 
 import cv2
@@ -450,8 +451,27 @@ def recording_clip(
     camera_name: str,
     start_ts: float,
     end_ts: float,
-    download: bool = False,
 ):
+    def run_download(ffmpeg_cmd: list[str], file_path: str):
+        with sp.Popen(
+            ffmpeg_cmd,
+            stderr=sp.PIPE,
+            stdout=sp.PIPE,
+            text=False,
+        ) as ffmpeg:
+            while True:
+                data = ffmpeg.stdout.read(8192)
+                if data is not None and len(data) > 0:
+                    yield data
+                else:
+                    if ffmpeg.returncode and ffmpeg.returncode != 0:
+                        logger.error(
+                            f"Failed to generate clip, ffmpeg logs: {ffmpeg.stderr.read()}"
+                        )
+                    else:
+                        FilePath(file_path).unlink(missing_ok=True)
+                    break
+
     recordings = (
         Recordings.select(
             Recordings.path,
@@ -467,18 +487,18 @@ def recording_clip(
         .order_by(Recordings.start_time.asc())
     )
 
-    playlist_lines = []
-    clip: Recordings
-    for clip in recordings:
-        playlist_lines.append(f"file '{clip.path}'")
-        # if this is the starting clip, add an inpoint
-        if clip.start_time < start_ts:
-            playlist_lines.append(f"inpoint {int(start_ts - clip.start_time)}")
-        # if this is the ending clip, add an outpoint
-        if clip.end_time > end_ts:
-            playlist_lines.append(f"outpoint {int(end_ts - clip.start_time)}")
-
-    file_name = sanitize_filename(f"clip_{camera_name}_{start_ts}-{end_ts}.mp4")
+    file_name = sanitize_filename(f"playlist_{camera_name}_{start_ts}-{end_ts}.txt")
+    file_path = f"/tmp/cache/{file_name}"
+    with open(file_path, "w") as file:
+        clip: Recordings
+        for clip in recordings:
+            file.write(f"file '{clip.path}'\n")
+            # if this is the starting clip, add an inpoint
+            if clip.start_time < start_ts:
+                file.write(f"inpoint {int(start_ts - clip.start_time)}\n")
+            # if this is the ending clip, add an outpoint
+            if clip.end_time > end_ts:
+                file.write(f"outpoint {int(end_ts - clip.start_time)}\n")
 
     if len(file_name) > 1000:
         return JSONResponse(
@@ -489,67 +509,32 @@ def recording_clip(
             status_code=403,
         )
 
-    path = os.path.join(CLIPS_DIR, f"cache/{file_name}")
-
     config: FrigateConfig = request.app.frigate_config
 
-    if not os.path.exists(path):
-        ffmpeg_cmd = [
-            config.ffmpeg.ffmpeg_path,
-            "-hide_banner",
-            "-y",
-            "-protocol_whitelist",
-            "pipe,file",
-            "-f",
-            "concat",
-            "-safe",
-            "0",
-            "-i",
-            "/dev/stdin",
-            "-c",
-            "copy",
-            "-movflags",
-            "+faststart",
-            path,
-        ]
-        p = sp.run(
-            ffmpeg_cmd,
-            input="\n".join(playlist_lines),
-            encoding="ascii",
-            capture_output=True,
-        )
+    ffmpeg_cmd = [
+        config.ffmpeg.ffmpeg_path,
+        "-hide_banner",
+        "-y",
+        "-protocol_whitelist",
+        "pipe,file",
+        "-f",
+        "concat",
+        "-safe",
+        "0",
+        "-i",
+        file_path,
+        "-c",
+        "copy",
+        "-movflags",
+        "frag_keyframe+empty_moov",
+        "-f",
+        "mp4",
+        "pipe:",
+    ]
 
-        if p.returncode != 0:
-            logger.error(p.stderr)
-            return JSONResponse(
-                content={
-                    "success": False,
-                    "message": "Could not create clip from recordings",
-                },
-                status_code=500,
-            )
-    else:
-        logger.debug(
-            f"Ignoring subsequent request for {path} as it already exists in the cache."
-        )
-
-    headers = {
-        "Content-Description": "File Transfer",
-        "Cache-Control": "no-cache",
-        "Content-Type": "video/mp4",
-        "Content-Length": str(os.path.getsize(path)),
-        # nginx: https://nginx.org/en/docs/http/ngx_http_proxy_module.html#proxy_ignore_headers
-        "X-Accel-Redirect": f"/clips/cache/{file_name}",
-    }
-
-    if download:
-        headers["Content-Disposition"] = "attachment; filename=%s" % file_name
-
-    return FileResponse(
-        path,
+    return StreamingResponse(
+        run_download(ffmpeg_cmd, file_path),
         media_type="video/mp4",
-        filename=file_name,
-        headers=headers,
     )
 
 
@@ -932,7 +917,7 @@ def grid_snapshot(
         ret, jpg = cv2.imencode(".jpg", frame, [int(cv2.IMWRITE_JPEG_QUALITY), 70])
 
         return Response(
-            jpg.tobytes,
+            jpg.tobytes(),
             media_type="image/jpeg",
             headers={"Cache-Control": "no-store"},
         )
@@ -1028,7 +1013,7 @@ def event_snapshot_clean(request: Request, event_id: str, download: bool = False
 
 
 @router.get("/events/{event_id}/clip.mp4")
-def event_clip(request: Request, event_id: str, download: bool = False):
+def event_clip(request: Request, event_id: str):
     try:
         event: Event = Event.get(Event.id == event_id)
     except DoesNotExist:
@@ -1048,7 +1033,7 @@ def event_clip(request: Request, event_id: str, download: bool = False):
         end_ts = (
             datetime.now().timestamp() if event.end_time is None else event.end_time
         )
-        return recording_clip(request, event.camera, event.start_time, end_ts, download)
+        return recording_clip(request, event.camera, event.start_time, end_ts)
 
     headers = {
         "Content-Description": "File Transfer",
@@ -1059,9 +1044,6 @@ def event_clip(request: Request, event_id: str, download: bool = False):
         "X-Accel-Redirect": f"/clips/{file_name}",
     }
 
-    if download:
-        headers["Content-Disposition"] = "attachment; filename=%s" % file_name
-
     return FileResponse(
         clip_path,
         media_type="video/mp4",
@@ -1471,7 +1453,6 @@ def preview_thumbnail(file_name: str):
 
     return Response(
         jpg_bytes,
-        # FIXME: Shouldn't it be either jpg or webp depending on the endpoint?
         media_type="image/webp",
         headers={
             "Content-Type": "image/webp",
@@ -1500,7 +1481,7 @@ def label_thumbnail(request: Request, camera_name: str, label: str):
         ret, jpg = cv2.imencode(".jpg", frame, [int(cv2.IMWRITE_JPEG_QUALITY), 70])
 
         return Response(
-            jpg.tobytes,
+            jpg.tobytes(),
             media_type="image/jpeg",
             headers={"Cache-Control": "no-store"},
         )
@@ -1546,13 +1527,13 @@ def label_snapshot(request: Request, camera_name: str, label: str):
         )
 
     try:
-        event = event_query.get()
-        return event_snapshot(request, event.id)
+        event: Event = event_query.get()
+        return event_snapshot(request, event.id, MediaEventsSnapshotQueryParams())
     except DoesNotExist:
         frame = np.zeros((720, 1280, 3), np.uint8)
-        ret, jpg = cv2.imencode(".jpg", frame, [int(cv2.IMWRITE_JPEG_QUALITY), 70])
+        _, jpg = cv2.imencode(".jpg", frame, [int(cv2.IMWRITE_JPEG_QUALITY), 70])
 
         return Response(
-            jpg.tobytes,
+            jpg.tobytes(),
             media_type="image/jpeg",
         )

frigate/api/review.py

@@ -12,11 +12,18 @@ from fastapi.responses import JSONResponse
 from peewee import Case, DoesNotExist, fn, operator
 from playhouse.shortcuts import model_to_dict
 
+from frigate.api.defs.generic_response import GenericResponse
+from frigate.api.defs.review_body import ReviewModifyMultipleBody
 from frigate.api.defs.review_query_parameters import (
     ReviewActivityMotionQueryParams,
     ReviewQueryParams,
     ReviewSummaryQueryParams,
 )
+from frigate.api.defs.review_responses import (
+    ReviewActivityMotionResponse,
+    ReviewSegmentResponse,
+    ReviewSummaryResponse,
+)
 from frigate.api.defs.tags import Tags
 from frigate.models import Recordings, ReviewSegment
 from frigate.util.builtin import get_tz_modifiers
@@ -26,7 +33,7 @@ logger = logging.getLogger(__name__)
 router = APIRouter(tags=[Tags.review])
 
 
-@router.get("/review")
+@router.get("/review", response_model=list[ReviewSegmentResponse])
 def review(params: ReviewQueryParams = Depends()):
     cameras = params.cameras
     labels = params.labels
@@ -102,7 +109,7 @@ def review(params: ReviewQueryParams = Depends()):
     return JSONResponse(content=[r for r in review])
 
 
-@router.get("/review/summary")
+@router.get("/review/summary", response_model=ReviewSummaryResponse)
 def review_summary(params: ReviewSummaryQueryParams = Depends()):
     hour_modifier, minute_modifier, seconds_offset = get_tz_modifiers(params.timezone)
     day_ago = (datetime.datetime.now() - datetime.timedelta(hours=24)).timestamp()
@@ -173,18 +180,6 @@ def review_summary(params: ReviewSummaryQueryParams = Depends()):
                     0,
                 )
             ).alias("reviewed_detection"),
-            fn.SUM(
-                Case(
-                    None,
-                    [
-                        (
-                            (ReviewSegment.severity == "significant_motion"),
-                            ReviewSegment.has_been_reviewed,
-                        )
-                    ],
-                    0,
-                )
-            ).alias("reviewed_motion"),
             fn.SUM(
                 Case(
                     None,
@@ -209,18 +204,6 @@ def review_summary(params: ReviewSummaryQueryParams = Depends()):
                     0,
                 )
             ).alias("total_detection"),
-            fn.SUM(
-                Case(
-                    None,
-                    [
-                        (
-                            (ReviewSegment.severity == "significant_motion"),
-                            1,
-                        )
-                    ],
-                    0,
-                )
-            ).alias("total_motion"),
         )
         .where(reduce(operator.and_, clauses))
         .dicts()
@@ -282,18 +265,6 @@ def review_summary(params: ReviewSummaryQueryParams = Depends()):
                     0,
                 )
             ).alias("reviewed_detection"),
-            fn.SUM(
-                Case(
-                    None,
-                    [
-                        (
-                            (ReviewSegment.severity == "significant_motion"),
-                            ReviewSegment.has_been_reviewed,
-                        )
-                    ],
-                    0,
-                )
-            ).alias("reviewed_motion"),
             fn.SUM(
                 Case(
                     None,
@@ -318,18 +289,6 @@ def review_summary(params: ReviewSummaryQueryParams = Depends()):
                     0,
                 )
             ).alias("total_detection"),
-            fn.SUM(
-                Case(
-                    None,
-                    [
-                        (
-                            (ReviewSegment.severity == "significant_motion"),
-                            1,
-                        )
-                    ],
-                    0,
-                )
-            ).alias("total_motion"),
         )
         .where(reduce(operator.and_, clauses))
         .group_by(
@@ -348,19 +307,10 @@ def review_summary(params: ReviewSummaryQueryParams = Depends()):
     return JSONResponse(content=data)
 
 
-@router.post("/reviews/viewed")
-def set_multiple_reviewed(body: dict = None):
-    json: dict[str, any] = body or {}
-    list_of_ids = json.get("ids", "")
-
-    if not list_of_ids or len(list_of_ids) == 0:
-        return JSONResponse(
-            context=({"success": False, "message": "Not a valid list of ids"}),
-            status_code=404,
-        )
-
+@router.post("/reviews/viewed", response_model=GenericResponse)
+def set_multiple_reviewed(body: ReviewModifyMultipleBody):
     ReviewSegment.update(has_been_reviewed=True).where(
-        ReviewSegment.id << list_of_ids
+        ReviewSegment.id << body.ids
     ).execute()
 
     return JSONResponse(
@@ -369,17 +319,9 @@ def set_multiple_reviewed(body: dict = None):
     )
 
 
-@router.post("/reviews/delete")
-def delete_reviews(body: dict = None):
-    json: dict[str, any] = body or {}
-    list_of_ids = json.get("ids", "")
-
-    if not list_of_ids or len(list_of_ids) == 0:
-        return JSONResponse(
-            content=({"success": False, "message": "Not a valid list of ids"}),
-            status_code=404,
-        )
-
+@router.post("/reviews/delete", response_model=GenericResponse)
+def delete_reviews(body: ReviewModifyMultipleBody):
+    list_of_ids = body.ids
     reviews = (
         ReviewSegment.select(
             ReviewSegment.camera,
@@ -424,7 +366,9 @@ def delete_reviews(body: dict = None):
     )
 
 
-@router.get("/review/activity/motion")
+@router.get(
+    "/review/activity/motion", response_model=list[ReviewActivityMotionResponse]
+)
 def motion_activity(params: ReviewActivityMotionQueryParams = Depends()):
     """Get motion and audio activity."""
     cameras = params.cameras
@@ -498,98 +442,44 @@ def motion_activity(params: ReviewActivityMotionQueryParams = Depends()):
     return JSONResponse(content=normalized)
 
 
-@router.get("/review/activity/audio")
-def audio_activity(params: ReviewActivityMotionQueryParams = Depends()):
-    """Get motion and audio activity."""
-    cameras = params.cameras
-    before = params.before or datetime.datetime.now().timestamp()
-    after = (
-        params.after
-        or (datetime.datetime.now() - datetime.timedelta(hours=1)).timestamp()
-    )
-    # get scale in seconds
-    scale = params.scale
-
-    clauses = [(Recordings.start_time > after) & (Recordings.end_time < before)]
-
-    if cameras != "all":
-        camera_list = cameras.split(",")
-        clauses.append((Recordings.camera << camera_list))
-
-    all_recordings: list[Recordings] = (
-        Recordings.select(
-            Recordings.start_time,
-            Recordings.duration,
-            Recordings.objects,
-            Recordings.dBFS,
-        )
-        .where(reduce(operator.and_, clauses))
-        .order_by(Recordings.start_time.asc())
-        .iterator()
-    )
-
-    # format is: { timestamp: segment_start_ts, motion: [0-100], audio: [0 - -100] }
-    # periods where active objects / audio was detected will cause audio to be scaled down
-    data: list[dict[str, float]] = []
-
-    for rec in all_recordings:
-        data.append(
-            {
-                "start_time": rec.start_time,
-                "audio": rec.dBFS if rec.objects == 0 else 0,
-            }
-        )
-
-    # resample data using pandas to get activity on scaled basis
-    df = pd.DataFrame(data, columns=["start_time", "audio"])
-    df = df.astype(dtype={"audio": "float16"})
-
-    # set date as datetime index
-    df["start_time"] = pd.to_datetime(df["start_time"], unit="s")
-    df.set_index(["start_time"], inplace=True)
-
-    # normalize data
-    df = df.resample(f"{scale}S").mean().fillna(0.0)
-    df["audio"] = (
-        (df["audio"] - df["audio"].max())
-        / (df["audio"].min() - df["audio"].max())
-        * -100
-    )
-
-    # change types for output
-    df.index = df.index.astype(int) // (10**9)
-    normalized = df.reset_index().to_dict("records")
-    return JSONResponse(content=normalized)
-
-
-@router.get("/review/event/{event_id}")
+@router.get("/review/event/{event_id}", response_model=ReviewSegmentResponse)
 def get_review_from_event(event_id: str):
     try:
-        return model_to_dict(
-            ReviewSegment.get(
-                ReviewSegment.data["detections"].cast("text") % f'*"{event_id}"*'
+        return JSONResponse(
+            model_to_dict(
+                ReviewSegment.get(
+                    ReviewSegment.data["detections"].cast("text") % f'*"{event_id}"*'
+                )
             )
         )
     except DoesNotExist:
-        return "Review item not found", 404
+        return JSONResponse(
+            content={"success": False, "message": "Review item not found"},
+            status_code=404,
+        )
 
 
-@router.get("/review/{event_id}")
-def get_review(event_id: str):
+@router.get("/review/{review_id}", response_model=ReviewSegmentResponse)
+def get_review(review_id: str):
     try:
-        return model_to_dict(ReviewSegment.get(ReviewSegment.id == event_id))
+        return JSONResponse(
+            content=model_to_dict(ReviewSegment.get(ReviewSegment.id == review_id))
+        )
     except DoesNotExist:
-        return "Review item not found", 404
+        return JSONResponse(
+            content={"success": False, "message": "Review item not found"},
+            status_code=404,
+        )
 
 
-@router.delete("/review/{event_id}/viewed")
-def set_not_reviewed(event_id: str):
+@router.delete("/review/{review_id}/viewed", response_model=GenericResponse)
+def set_not_reviewed(review_id: str):
     try:
-        review: ReviewSegment = ReviewSegment.get(ReviewSegment.id == event_id)
+        review: ReviewSegment = ReviewSegment.get(ReviewSegment.id == review_id)
     except DoesNotExist:
         return JSONResponse(
             content=(
-                {"success": False, "message": "Review " + event_id + " not found"}
+                {"success": False, "message": "Review " + review_id + " not found"}
             ),
             status_code=404,
         )
@@ -598,6 +488,8 @@ def set_not_reviewed(event_id: str):
     review.save()
 
     return JSONResponse(
-        content=({"success": True, "message": "Reviewed " + event_id + " not viewed"}),
+        content=(
+            {"success": True, "message": "Set Review " + review_id + " as not viewed"}
+        ),
         status_code=200,
     )

frigate/app.py

@@ -68,6 +68,7 @@ from frigate.stats.util import stats_init
 from frigate.storage import StorageMaintainer
 from frigate.timeline import TimelineProcessor
 from frigate.util.builtin import empty_and_close_queue
+from frigate.util.image import SharedMemoryFrameManager, UntrackedSharedMemory
 from frigate.util.object import get_camera_regions_grid
 from frigate.version import VERSION
 from frigate.video import capture_camera, track_camera
@@ -90,6 +91,7 @@ class FrigateApp:
         self.processes: dict[str, int] = {}
         self.embeddings: Optional[EmbeddingsContext] = None
         self.region_grids: dict[str, list[list[dict[str, int]]]] = {}
+        self.frame_manager = SharedMemoryFrameManager()
         self.config = config
 
     def ensure_dirs(self) -> None:
@@ -325,20 +327,20 @@ class FrigateApp:
                         for det in self.config.detectors.values()
                     ]
                 )
-                shm_in = mp.shared_memory.SharedMemory(
+                shm_in = UntrackedSharedMemory(
                     name=name,
                     create=True,
                     size=largest_frame,
                 )
             except FileExistsError:
-                shm_in = mp.shared_memory.SharedMemory(name=name)
+                shm_in = UntrackedSharedMemory(name=name)
 
             try:
-                shm_out = mp.shared_memory.SharedMemory(
+                shm_out = UntrackedSharedMemory(
                     name=f"out-{name}", create=True, size=20 * 6 * 4
                 )
             except FileExistsError:
-                shm_out = mp.shared_memory.SharedMemory(name=f"out-{name}")
+                shm_out = UntrackedSharedMemory(name=f"out-{name}")
 
             self.detection_shms.append(shm_in)
             self.detection_shms.append(shm_out)
@@ -431,6 +433,11 @@ class FrigateApp:
                 logger.info(f"Capture process not started for disabled camera {name}")
                 continue
 
+            # pre-create shms
+            for i in range(shm_frame_count):
+                frame_size = config.frame_shape_yuv[0] * config.frame_shape_yuv[1]
+                self.frame_manager.create(f"{config.name}{i}", frame_size)
+
             capture_process = util.Process(
                 target=capture_camera,
                 name=f"camera_capture:{name}",
@@ -513,15 +520,18 @@ class FrigateApp:
                     1,
                 )
 
-        shm_frame_count = min(50, int(available_shm / (cam_total_frame_size)))
+        if cam_total_frame_size == 0.0:
+            return 0
+
+        shm_frame_count = min(200, int(available_shm / (cam_total_frame_size)))
 
         logger.debug(
             f"Calculated total camera size {available_shm} / {cam_total_frame_size} :: {shm_frame_count} frames for each camera in SHM"
         )
 
-        if shm_frame_count < 10:
+        if shm_frame_count < 20:
             logger.warning(
-                f"The current SHM size of {total_shm}MB is too small, recommend increasing it to at least {round(min_req_shm + cam_total_frame_size * 10)}MB."
+                f"The current SHM size of {total_shm}MB is too small, recommend increasing it to at least {round(min_req_shm + cam_total_frame_size * 20)}MB."
             )
 
         return shm_frame_count
@@ -581,12 +591,12 @@ class FrigateApp:
         self.init_recording_manager()
         self.init_review_segment_manager()
         self.init_go2rtc()
+        self.start_detectors()
+        self.init_embeddings_manager()
         self.bind_database()
         self.check_db_data_migrations()
         self.init_inter_process_communicator()
         self.init_dispatcher()
-        self.start_detectors()
-        self.init_embeddings_manager()
         self.init_embeddings_client()
         self.start_video_output_processor()
         self.start_ptz_autotracker()
@@ -699,7 +709,7 @@ class FrigateApp:
 
         # Save embeddings stats to disk
         if self.embeddings:
-            self.embeddings.save_stats()
+            self.embeddings.stop()
 
         # Stop Communicators
         self.inter_process_communicator.stop()
@@ -707,6 +717,7 @@ class FrigateApp:
         self.event_metadata_updater.stop()
         self.inter_zmq_proxy.stop()
 
+        self.frame_manager.cleanup()
         while len(self.detection_shms) > 0:
             shm = self.detection_shms.pop()
             shm.close()

frigate/comms/dispatcher.py

@@ -15,13 +15,14 @@ from frigate.const import (
     INSERT_PREVIEW,
     REQUEST_REGION_GRID,
     UPDATE_CAMERA_ACTIVITY,
+    UPDATE_EMBEDDINGS_REINDEX_PROGRESS,
     UPDATE_EVENT_DESCRIPTION,
     UPDATE_MODEL_STATE,
     UPSERT_REVIEW_SEGMENT,
 )
 from frigate.models import Event, Previews, Recordings, ReviewSegment
 from frigate.ptz.onvif import OnvifCommandEnum, OnvifController
-from frigate.types import ModelStatusTypesEnum
+from frigate.types import ModelStatusTypesEnum, TrackedObjectUpdateTypesEnum
 from frigate.util.object import get_camera_regions_grid
 from frigate.util.services import restart_frigate
 
@@ -63,6 +64,9 @@ class Dispatcher:
         self.onvif = onvif
         self.ptz_metrics = ptz_metrics
         self.comms = communicators
+        self.camera_activity = {}
+        self.model_state = {}
+        self.embeddings_reindex = {}
 
         self._camera_settings_handlers: dict[str, Callable] = {
             "audio": self._on_audio_command,
@@ -84,37 +88,25 @@ class Dispatcher:
         for comm in self.comms:
             comm.subscribe(self._receive)
 
-        self.camera_activity = {}
-        self.model_state = {}
-
     def _receive(self, topic: str, payload: str) -> Optional[Any]:
         """Handle receiving of payload from communicators."""
-        if topic.endswith("set"):
+
+        def handle_camera_command(command_type, camera_name, command, payload):
             try:
-                # example /cam_name/detect/set payload=ON|OFF
-                if topic.count("/") == 2:
-                    camera_name = topic.split("/")[-3]
-                    command = topic.split("/")[-2]
+                if command_type == "set":
                     self._camera_settings_handlers[command](camera_name, payload)
-                elif topic.count("/") == 1:
-                    command = topic.split("/")[-2]
-                    self._global_settings_handlers[command](payload)
-            except IndexError:
-                logger.error(f"Received invalid set command: {topic}")
-                return
-        elif topic.endswith("ptz"):
-            try:
-                # example /cam_name/ptz payload=MOVE_UP|MOVE_DOWN|STOP...
-                camera_name = topic.split("/")[-2]
-                self._on_ptz_command(camera_name, payload)
-            except IndexError:
-                logger.error(f"Received invalid ptz command: {topic}")
-                return
-        elif topic == "restart":
+                elif command_type == "ptz":
+                    self._on_ptz_command(camera_name, payload)
+            except KeyError:
+                logger.error(f"Invalid command type or handler: {command_type}")
+
+        def handle_restart():
             restart_frigate()
-        elif topic == INSERT_MANY_RECORDINGS:
+
+        def handle_insert_many_recordings():
             Recordings.insert_many(payload).execute()
-        elif topic == REQUEST_REGION_GRID:
+
+        def handle_request_region_grid():
             camera = payload
             grid = get_camera_regions_grid(
                 camera,
@@ -122,40 +114,63 @@ class Dispatcher:
                 max(self.config.model.width, self.config.model.height),
             )
             return grid
-        elif topic == INSERT_PREVIEW:
+
+        def handle_insert_preview():
             Previews.insert(payload).execute()
-        elif topic == UPSERT_REVIEW_SEGMENT:
-            (
-                ReviewSegment.insert(payload)
-                .on_conflict(
-                    conflict_target=[ReviewSegment.id],
-                    update=payload,
-                )
-                .execute()
-            )
-        elif topic == CLEAR_ONGOING_REVIEW_SEGMENTS:
-            ReviewSegment.update(end_time=datetime.datetime.now().timestamp()).where(
-                ReviewSegment.end_time == None
+
+        def handle_upsert_review_segment():
+            ReviewSegment.insert(payload).on_conflict(
+                conflict_target=[ReviewSegment.id],
+                update=payload,
             ).execute()
-        elif topic == UPDATE_CAMERA_ACTIVITY:
+
+        def handle_clear_ongoing_review_segments():
+            ReviewSegment.update(end_time=datetime.datetime.now().timestamp()).where(
+                ReviewSegment.end_time.is_null(True)
+            ).execute()
+
+        def handle_update_camera_activity():
             self.camera_activity = payload
-        elif topic == UPDATE_EVENT_DESCRIPTION:
+
+        def handle_update_event_description():
             event: Event = Event.get(Event.id == payload["id"])
             event.data["description"] = payload["description"]
             event.save()
             self.publish(
-                "event_update",
-                json.dumps({"id": event.id, "description": event.data["description"]}),
+                "tracked_object_update",
+                json.dumps(
+                    {
+                        "type": TrackedObjectUpdateTypesEnum.description,
+                        "id": event.id,
+                        "description": event.data["description"],
+                    }
+                ),
             )
-        elif topic == UPDATE_MODEL_STATE:
-            model = payload["model"]
-            state = payload["state"]
-            self.model_state[model] = ModelStatusTypesEnum[state]
-            self.publish("model_state", json.dumps(self.model_state))
-        elif topic == "modelState":
-            model_state = self.model_state.copy()
-            self.publish("model_state", json.dumps(model_state))
-        elif topic == "onConnect":
+
+        def handle_update_model_state():
+            if payload:
+                model = payload["model"]
+                state = payload["state"]
+                self.model_state[model] = ModelStatusTypesEnum[state]
+                self.publish("model_state", json.dumps(self.model_state))
+
+        def handle_model_state():
+            self.publish("model_state", json.dumps(self.model_state.copy()))
+
+        def handle_update_embeddings_reindex_progress():
+            self.embeddings_reindex = payload
+            self.publish(
+                "embeddings_reindex_progress",
+                json.dumps(payload),
+            )
+
+        def handle_embeddings_reindex_progress():
+            self.publish(
+                "embeddings_reindex_progress",
+                json.dumps(self.embeddings_reindex.copy()),
+            )
+
+        def handle_on_connect():
             camera_status = self.camera_activity.copy()
 
             for camera in camera_status.keys():
@@ -170,6 +185,51 @@ class Dispatcher:
                 }
 
             self.publish("camera_activity", json.dumps(camera_status))
+            self.publish("model_state", json.dumps(self.model_state.copy()))
+            self.publish(
+                "embeddings_reindex_progress",
+                json.dumps(self.embeddings_reindex.copy()),
+            )
+
+        # Dictionary mapping topic to handlers
+        topic_handlers = {
+            INSERT_MANY_RECORDINGS: handle_insert_many_recordings,
+            REQUEST_REGION_GRID: handle_request_region_grid,
+            INSERT_PREVIEW: handle_insert_preview,
+            UPSERT_REVIEW_SEGMENT: handle_upsert_review_segment,
+            CLEAR_ONGOING_REVIEW_SEGMENTS: handle_clear_ongoing_review_segments,
+            UPDATE_CAMERA_ACTIVITY: handle_update_camera_activity,
+            UPDATE_EVENT_DESCRIPTION: handle_update_event_description,
+            UPDATE_MODEL_STATE: handle_update_model_state,
+            UPDATE_EMBEDDINGS_REINDEX_PROGRESS: handle_update_embeddings_reindex_progress,
+            "restart": handle_restart,
+            "embeddingsReindexProgress": handle_embeddings_reindex_progress,
+            "modelState": handle_model_state,
+            "onConnect": handle_on_connect,
+        }
+
+        if topic.endswith("set") or topic.endswith("ptz"):
+            try:
+                parts = topic.split("/")
+                if len(parts) == 3 and topic.endswith("set"):
+                    # example /cam_name/detect/set payload=ON|OFF
+                    camera_name = parts[-3]
+                    command = parts[-2]
+                    handle_camera_command("set", camera_name, command, payload)
+                elif len(parts) == 2 and topic.endswith("set"):
+                    command = parts[-2]
+                    self._global_settings_handlers[command](payload)
+                elif len(parts) == 2 and topic.endswith("ptz"):
+                    # example /cam_name/ptz payload=MOVE_UP|MOVE_DOWN|STOP...
+                    camera_name = parts[-2]
+                    handle_camera_command("ptz", camera_name, "", payload)
+            except IndexError:
+                logger.error(
+                    f"Received invalid {topic.split('/')[-1]} command: {topic}"
+                )
+                return
+        elif topic in topic_handlers:
+            return topic_handlers[topic]()
         else:
             self.publish(topic, payload, retain=False)
 

frigate/comms/embeddings_updater.py

@@ -0,0 +1,65 @@
+"""Facilitates communication between processes."""
+
+from enum import Enum
+from typing import Callable
+
+import zmq
+
+SOCKET_REP_REQ = "ipc:///tmp/cache/embeddings"
+
+
+class EmbeddingsRequestEnum(Enum):
+    embed_description = "embed_description"
+    embed_thumbnail = "embed_thumbnail"
+    generate_search = "generate_search"
+
+
+class EmbeddingsResponder:
+    def __init__(self) -> None:
+        self.context = zmq.Context()
+        self.socket = self.context.socket(zmq.REP)
+        self.socket.bind(SOCKET_REP_REQ)
+
+    def check_for_request(self, process: Callable) -> None:
+        while True:  # load all messages that are queued
+            has_message, _, _ = zmq.select([self.socket], [], [], 0.1)
+
+            if not has_message:
+                break
+
+            try:
+                (topic, value) = self.socket.recv_json(flags=zmq.NOBLOCK)
+
+                response = process(topic, value)
+
+                if response is not None:
+                    self.socket.send_json(response)
+                else:
+                    self.socket.send_json([])
+            except zmq.ZMQError:
+                break
+
+    def stop(self) -> None:
+        self.socket.close()
+        self.context.destroy()
+
+
+class EmbeddingsRequestor:
+    """Simplifies sending data to EmbeddingsResponder and getting a reply."""
+
+    def __init__(self) -> None:
+        self.context = zmq.Context()
+        self.socket = self.context.socket(zmq.REQ)
+        self.socket.connect(SOCKET_REP_REQ)
+
+    def send_data(self, topic: str, data: any) -> str:
+        """Sends data and then waits for reply."""
+        try:
+            self.socket.send_json((topic, data))
+            return self.socket.recv_json()
+        except zmq.ZMQError:
+            return ""
+
+    def stop(self) -> None:
+        self.socket.close()
+        self.context.destroy()

frigate/comms/event_metadata_updater.py

@@ -39,7 +39,7 @@ class EventMetadataSubscriber(Subscriber):
         super().__init__(topic)
 
     def check_for_update(
-        self, timeout: float = None
+        self, timeout: float = 1
     ) -> Optional[tuple[EventMetadataTypeEnum, str, RegenerateDescriptionEnum]]:
         return super().check_for_update(timeout)
 

frigate/comms/events_updater.py

@@ -14,7 +14,7 @@ class EventUpdatePublisher(Publisher):
         super().__init__("update")
 
     def publish(
-        self, payload: tuple[EventTypeEnum, EventStateEnum, str, dict[str, any]]
+        self, payload: tuple[EventTypeEnum, EventStateEnum, str, str, dict[str, any]]
     ) -> None:
         super().publish(payload)
 

frigate/comms/inter_process.py

@@ -65,8 +65,11 @@ class InterProcessRequestor:
 
     def send_data(self, topic: str, data: any) -> any:
         """Sends data and then waits for reply."""
-        self.socket.send_json((topic, data))
-        return self.socket.recv_json()
+        try:
+            self.socket.send_json((topic, data))
+            return self.socket.recv_json()
+        except zmq.ZMQError:
+            return ""
 
     def stop(self) -> None:
         self.socket.close()

frigate/comms/mqtt.py

@@ -17,7 +17,7 @@ class MqttClient(Communicator):  # type: ignore[misc]
     def __init__(self, config: FrigateConfig) -> None:
         self.config = config
         self.mqtt_config = config.mqtt
-        self.connected: bool = False
+        self.connected = False
 
     def subscribe(self, receiver: Callable) -> None:
         """Wrapper for allowing dispatcher to subscribe."""
@@ -27,7 +27,7 @@ class MqttClient(Communicator):  # type: ignore[misc]
     def publish(self, topic: str, payload: Any, retain: bool = False) -> None:
         """Wrapper for publishing when client is in valid state."""
         if not self.connected:
-            logger.error(f"Unable to publish to {topic}: client is not connected")
+            logger.debug(f"Unable to publish to {topic}: client is not connected")
             return
 
         self.client.publish(
@@ -133,7 +133,7 @@ class MqttClient(Communicator):  # type: ignore[misc]
         """Mqtt connection callback."""
         threading.current_thread().name = "mqtt"
         if reason_code != 0:
-            if reason_code == "Server Unavailable":
+            if reason_code == "Server unavailable":
                 logger.error(
                     "Unable to connect to MQTT server: MQTT Server unavailable"
                 )
@@ -173,6 +173,7 @@ class MqttClient(Communicator):  # type: ignore[misc]
             client_id=self.mqtt_config.client_id,
         )
         self.client.on_connect = self._on_connect
+        self.client.on_disconnect = self._on_disconnect
         self.client.will_set(
             self.mqtt_config.topic_prefix + "/available",
             payload="offline",
@@ -197,14 +198,6 @@ class MqttClient(Communicator):  # type: ignore[misc]
 
         for name in self.config.cameras.keys():
             for callback in callback_types:
-                # We need to pre-clear existing set topics because in previous
-                # versions the webUI retained on the /set topic but this is
-                # no longer the case.
-                self.client.publish(
-                    f"{self.mqtt_config.topic_prefix}/{name}/{callback}/set",
-                    None,
-                    retain=True,
-                )
                 self.client.message_callback_add(
                     f"{self.mqtt_config.topic_prefix}/{name}/{callback}/set",
                     self.on_mqtt_command,

frigate/config/auth.py

@@ -13,7 +13,7 @@ class AuthConfig(FrigateBaseModel):
         default=False, title="Reset the admin password on startup"
     )
     cookie_name: str = Field(
-        default="frigate_token", title="Name for jwt token cookie", pattern=r"^[a-z]_*$"
+        default="frigate_token", title="Name for jwt token cookie", pattern=r"^[a-z_]+$"
     )
     cookie_secure: bool = Field(default=False, title="Set secure flag on cookie")
     session_length: int = Field(

frigate/config/camera/genai.py

@@ -23,7 +23,7 @@ class GenAICameraConfig(BaseModel):
         default=False, title="Use snapshots for generating descriptions."
     )
     prompt: str = Field(
-        default="Describe the {label} in the sequence of images with as much detail as possible. Do not describe the background.",
+        default="Analyze the sequence of images containing the {label}. Focus on the likely intent or behavior of the {label} based on its actions and movement, rather than describing its appearance or the surroundings. Consider what the {label} is doing, why, and what it might do next.",
         title="Default caption prompt.",
     )
     object_prompts: dict[str, str] = Field(
@@ -51,7 +51,7 @@ class GenAICameraConfig(BaseModel):
 class GenAIConfig(FrigateBaseModel):
     enabled: bool = Field(default=False, title="Enable GenAI.")
     prompt: str = Field(
-        default="Describe the {label} in the sequence of images with as much detail as possible. Do not describe the background.",
+        default="Analyze the sequence of images containing the {label}. Focus on the likely intent or behavior of the {label} based on its actions and movement, rather than describing its appearance or the surroundings. Consider what the {label} is doing, why, and what it might do next.",
         title="Default caption prompt.",
     )
     object_prompts: dict[str, str] = Field(

frigate/config/camera/record.py

@@ -94,3 +94,10 @@ class RecordConfig(FrigateBaseModel):
     enabled_in_config: Optional[bool] = Field(
         default=None, title="Keep track of original state of recording."
     )
+
+    @property
+    def event_pre_capture(self) -> int:
+        return max(
+            self.alerts.pre_capture,
+            self.detections.pre_capture,
+        )

frigate/config/config.py

@@ -67,7 +67,7 @@ logger = logging.getLogger(__name__)
 
 yaml = YAML()
 
-DEFAULT_CONFIG_FILES = ["/config/config.yaml", "/config/config.yml"]
+DEFAULT_CONFIG_FILE = "/config/config.yml"
 DEFAULT_CONFIG = """
 mqtt:
   enabled: False
@@ -634,27 +634,23 @@ class FrigateConfig(FrigateBaseModel):
 
     @classmethod
     def load(cls, **kwargs):
-        config_path = os.environ.get("CONFIG_FILE")
+        config_path = os.environ.get("CONFIG_FILE", DEFAULT_CONFIG_FILE)
 
-        # No explicit configuration file, try to find one in the default paths.
-        if config_path is None:
-            for path in DEFAULT_CONFIG_FILES:
-                if os.path.isfile(path):
-                    config_path = path
-                    break
+        if not os.path.isfile(config_path):
+            config_path = config_path.replace("yml", "yaml")
 
         # No configuration file found, create one.
         new_config = False
-        if config_path is None:
+        if not os.path.isfile(config_path):
             logger.info("No config file found, saving default config")
-            config_path = DEFAULT_CONFIG_FILES[-1]
+            config_path = DEFAULT_CONFIG_FILE
             new_config = True
         else:
             # Check if the config file needs to be migrated.
             migrate_frigate_config(config_path)
 
         # Finally, load the resulting configuration file.
-        with open(config_path, "a+") as f:
+        with open(config_path, "a+" if new_config else "r") as f:
             # Only write the default config if the opened file is non-empty. This can happen as
             # a race condition. It's extremely unlikely, but eh. Might as well check it.
             if new_config and f.tell() == 0:

frigate/config/env.py

@@ -23,7 +23,7 @@ EnvString = Annotated[str, AfterValidator(validate_env_string)]
 
 def validate_env_vars(v: dict[str, str], info: ValidationInfo) -> dict[str, str]:
     if isinstance(info.context, dict) and info.context.get("install", False):
-        for k, v in v:
+        for k, v in v.items():
             os.environ[k] = v
 
     return v

frigate/config/semantic_search.py

@@ -12,3 +12,6 @@ class SemanticSearchConfig(FrigateBaseModel):
     reindex: Optional[bool] = Field(
         default=False, title="Reindex all detections on startup."
     )
+    model_size: str = Field(
+        default="small", title="The size of the embeddings model used."
+    )

frigate/const.py

@@ -17,7 +17,21 @@ PLUS_API_HOST = "https://api.frigate.video"
 
 DEFAULT_ATTRIBUTE_LABEL_MAP = {
     "person": ["amazon", "face"],
-    "car": ["amazon", "fedex", "license_plate", "ups"],
+    "car": [
+        "amazon",
+        "an_post",
+        "dhl",
+        "dpd",
+        "fedex",
+        "gls",
+        "license_plate",
+        "nzpost",
+        "postnl",
+        "postnord",
+        "purolator",
+        "ups",
+        "usps",
+    ],
 }
 LABEL_CONSOLIDATION_MAP = {
     "car": 0.8,
@@ -85,6 +99,7 @@ CLEAR_ONGOING_REVIEW_SEGMENTS = "clear_ongoing_review_segments"
 UPDATE_CAMERA_ACTIVITY = "update_camera_activity"
 UPDATE_EVENT_DESCRIPTION = "update_event_description"
 UPDATE_MODEL_STATE = "update_model_state"
+UPDATE_EMBEDDINGS_REINDEX_PROGRESS = "handle_embeddings_reindex_progress"
 
 # Stats Values
 

frigate/db/sqlitevecq.py

@@ -20,3 +20,34 @@ class SqliteVecQueueDatabase(SqliteQueueDatabase):
         conn.enable_load_extension(True)
         conn.load_extension(self.sqlite_vec_path)
         conn.enable_load_extension(False)
+
+    def delete_embeddings_thumbnail(self, event_ids: list[str]) -> None:
+        ids = ",".join(["?" for _ in event_ids])
+        self.execute_sql(f"DELETE FROM vec_thumbnails WHERE id IN ({ids})", event_ids)
+
+    def delete_embeddings_description(self, event_ids: list[str]) -> None:
+        ids = ",".join(["?" for _ in event_ids])
+        self.execute_sql(f"DELETE FROM vec_descriptions WHERE id IN ({ids})", event_ids)
+
+    def drop_embeddings_tables(self) -> None:
+        self.execute_sql("""
+            DROP TABLE vec_descriptions;
+        """)
+        self.execute_sql("""
+            DROP TABLE vec_thumbnails;
+        """)
+
+    def create_embeddings_tables(self) -> None:
+        """Create vec0 virtual table for embeddings"""
+        self.execute_sql("""
+            CREATE VIRTUAL TABLE IF NOT EXISTS vec_thumbnails USING vec0(
+                id TEXT PRIMARY KEY,
+                thumbnail_embedding FLOAT[768] distance_metric=cosine
+            );
+        """)
+        self.execute_sql("""
+            CREATE VIRTUAL TABLE IF NOT EXISTS vec_descriptions USING vec0(
+                id TEXT PRIMARY KEY,
+                description_embedding FLOAT[768] distance_metric=cosine
+            );
+        """)

frigate/detectors/detector_config.py

@@ -27,6 +27,11 @@ class InputTensorEnum(str, Enum):
     nhwc = "nhwc"
 
 
+class InputDTypeEnum(str, Enum):
+    float = "float"
+    int = "int"
+
+
 class ModelTypeEnum(str, Enum):
     ssd = "ssd"
     yolox = "yolox"
@@ -53,12 +58,16 @@ class ModelConfig(BaseModel):
     input_pixel_format: PixelFormatEnum = Field(
         default=PixelFormatEnum.rgb, title="Model Input Pixel Color Format"
     )
+    input_dtype: InputDTypeEnum = Field(
+        default=InputDTypeEnum.int, title="Model Input D Type"
+    )
     model_type: ModelTypeEnum = Field(
         default=ModelTypeEnum.ssd, title="Object Detection Model Type"
     )
     _merged_labelmap: Optional[Dict[int, str]] = PrivateAttr()
     _colormap: Dict[int, Tuple[int, int, int]] = PrivateAttr()
     _all_attributes: list[str] = PrivateAttr()
+    _all_attribute_logos: list[str] = PrivateAttr()
     _model_hash: str = PrivateAttr()
 
     @property
@@ -73,6 +82,10 @@ class ModelConfig(BaseModel):
     def all_attributes(self) -> list[str]:
         return self._all_attributes
 
+    @property
+    def all_attribute_logos(self) -> list[str]:
+        return self._all_attribute_logos
+
     @property
     def model_hash(self) -> str:
         return self._model_hash
@@ -93,6 +106,9 @@ class ModelConfig(BaseModel):
             unique_attributes.update(attributes)
 
         self._all_attributes = list(unique_attributes)
+        self._all_attribute_logos = list(
+            unique_attributes - set(["face", "license_plate"])
+        )
 
     def check_and_load_plus_model(
         self, plus_api: PlusApi, detector: str = None
@@ -140,6 +156,9 @@ class ModelConfig(BaseModel):
             unique_attributes.update(attributes)
 
         self._all_attributes = list(unique_attributes)
+        self._all_attribute_logos = list(
+            unique_attributes - set(["face", "license_plate"])
+        )
 
         self._merged_labelmap = {
             **{int(key): val for key, val in model_info["labelMap"].items()},
@@ -157,10 +176,14 @@ class ModelConfig(BaseModel):
             self._model_hash = file_hash.hexdigest()
 
     def create_colormap(self, enabled_labels: set[str]) -> None:
-        """Get a list of colors for enabled labels."""
-        colors = generate_color_palette(len(enabled_labels))
-
-        self._colormap = {label: color for label, color in zip(enabled_labels, colors)}
+        """Get a list of colors for enabled labels that aren't attributes."""
+        enabled_trackable_labels = list(
+            filter(lambda label: label not in self._all_attributes, enabled_labels)
+        )
+        colors = generate_color_palette(len(enabled_trackable_labels))
+        self._colormap = {
+            label: color for label, color in zip(enabled_trackable_labels, colors)
+        }
 
     model_config = ConfigDict(extra="forbid", protected_namespaces=())
 

frigate/detectors/plugins/onnx.py

@@ -54,7 +54,7 @@ class ONNXDetector(DetectionApi):
 
         logger.info(f"ONNX: {path} loaded")
 
-    def detect_raw(self, tensor_input):
+    def detect_raw(self, tensor_input: np.ndarray):
         model_input_name = self.model.get_inputs()[0].name
         tensor_output = self.model.run(None, {model_input_name: tensor_input})
 

frigate/detectors/plugins/openvino.py

@@ -3,6 +3,7 @@ import os
 
 import numpy as np
 import openvino as ov
+import openvino.properties as props
 from pydantic import Field
 from typing_extensions import Literal
 
@@ -34,6 +35,8 @@ class OvDetector(DetectionApi):
             logger.error(f"OpenVino model file {detector_config.model.path} not found.")
             raise FileNotFoundError
 
+        os.makedirs("/config/model_cache/openvino", exist_ok=True)
+        self.ov_core.set_property({props.cache_dir: "/config/model_cache/openvino"})
         self.interpreter = self.ov_core.compile_model(
             model=detector_config.model.path, device_name=detector_config.device
         )

frigate/detectors/plugins/rocm.py

@@ -98,9 +98,7 @@ class ROCmDetector(DetectionApi):
         else:
             logger.info(f"AMD/ROCm: loading model from {path}")
 
-            if path.endswith(".onnx"):
-                self.model = migraphx.parse_onnx(path)
-            elif (
+            if (
                 path.endswith(".tf")
                 or path.endswith(".tf2")
                 or path.endswith(".tflite")
@@ -108,7 +106,7 @@ class ROCmDetector(DetectionApi):
                 # untested
                 self.model = migraphx.parse_tf(path)
             else:
-                raise Exception(f"AMD/ROCm: unknown model format {path}")
+                self.model = migraphx.parse_onnx(path)
 
             logger.info("AMD/ROCm: compiling the model")
 

frigate/embeddings/__init__.py

@@ -7,17 +7,18 @@ import os
 import signal
 import threading
 from types import FrameType
-from typing import Optional
+from typing import Optional, Union
 
 from setproctitle import setproctitle
 
+from frigate.comms.embeddings_updater import EmbeddingsRequestEnum, EmbeddingsRequestor
 from frigate.config import FrigateConfig
 from frigate.const import CONFIG_DIR
 from frigate.db.sqlitevecq import SqliteVecQueueDatabase
 from frigate.models import Event
+from frigate.util.builtin import serialize
 from frigate.util.services import listen
 
-from .embeddings import Embeddings
 from .maintainer import EmbeddingMaintainer
 from .util import ZScoreNormalization
 
@@ -55,12 +56,6 @@ def manage_embeddings(config: FrigateConfig) -> None:
     models = [Event]
     db.bind(models)
 
-    embeddings = Embeddings(db)
-
-    # Check if we need to re-index events
-    if config.semantic_search.reindex:
-        embeddings.reindex()
-
     maintainer = EmbeddingMaintainer(
         db,
         config,
@@ -71,9 +66,10 @@ def manage_embeddings(config: FrigateConfig) -> None:
 
 class EmbeddingsContext:
     def __init__(self, db: SqliteVecQueueDatabase):
-        self.embeddings = Embeddings(db)
+        self.db = db
         self.thumb_stats = ZScoreNormalization()
-        self.desc_stats = ZScoreNormalization(scale_factor=3, bias=-2.5)
+        self.desc_stats = ZScoreNormalization()
+        self.requestor = EmbeddingsRequestor()
 
         # load stats from disk
         try:
@@ -84,7 +80,7 @@ class EmbeddingsContext:
         except FileNotFoundError:
             pass
 
-    def save_stats(self):
+    def stop(self):
         """Write the stats to disk as JSON on exit."""
         contents = {
             "thumb_stats": self.thumb_stats.to_dict(),
@@ -92,3 +88,109 @@ class EmbeddingsContext:
         }
         with open(os.path.join(CONFIG_DIR, ".search_stats.json"), "w") as f:
             json.dump(contents, f)
+        self.requestor.stop()
+
+    def search_thumbnail(
+        self, query: Union[Event, str], event_ids: list[str] = None
+    ) -> list[tuple[str, float]]:
+        if query.__class__ == Event:
+            cursor = self.db.execute_sql(
+                """
+                SELECT thumbnail_embedding FROM vec_thumbnails WHERE id = ?
+                """,
+                [query.id],
+            )
+
+            row = cursor.fetchone() if cursor else None
+
+            if row:
+                query_embedding = row[0]
+            else:
+                # If no embedding found, generate it and return it
+                data = self.requestor.send_data(
+                    EmbeddingsRequestEnum.embed_thumbnail.value,
+                    {"id": str(query.id), "thumbnail": str(query.thumbnail)},
+                )
+
+                if not data:
+                    return []
+
+                query_embedding = serialize(data)
+        else:
+            data = self.requestor.send_data(
+                EmbeddingsRequestEnum.generate_search.value, query
+            )
+
+            if not data:
+                return []
+
+            query_embedding = serialize(data)
+
+        sql_query = """
+            SELECT
+                id,
+                distance
+            FROM vec_thumbnails
+            WHERE thumbnail_embedding MATCH ?
+                AND k = 100
+        """
+
+        # Add the IN clause if event_ids is provided and not empty
+        # this is the only filter supported by sqlite-vec as of 0.1.3
+        # but it seems to be broken in this version
+        if event_ids:
+            sql_query += " AND id IN ({})".format(",".join("?" * len(event_ids)))
+
+        # order by distance DESC is not implemented in this version of sqlite-vec
+        # when it's implemented, we can use cosine similarity
+        sql_query += " ORDER BY distance"
+
+        parameters = [query_embedding] + event_ids if event_ids else [query_embedding]
+
+        results = self.db.execute_sql(sql_query, parameters).fetchall()
+
+        return results
+
+    def search_description(
+        self, query_text: str, event_ids: list[str] = None
+    ) -> list[tuple[str, float]]:
+        data = self.requestor.send_data(
+            EmbeddingsRequestEnum.generate_search.value, query_text
+        )
+
+        if not data:
+            return []
+
+        query_embedding = serialize(data)
+
+        # Prepare the base SQL query
+        sql_query = """
+            SELECT
+                id,
+                distance
+            FROM vec_descriptions
+            WHERE description_embedding MATCH ?
+                AND k = 100
+        """
+
+        # Add the IN clause if event_ids is provided and not empty
+        # this is the only filter supported by sqlite-vec as of 0.1.3
+        # but it seems to be broken in this version
+        if event_ids:
+            sql_query += " AND id IN ({})".format(",".join("?" * len(event_ids)))
+
+        # order by distance DESC is not implemented in this version of sqlite-vec
+        # when it's implemented, we can use cosine similarity
+        sql_query += " ORDER BY distance"
+
+        parameters = [query_embedding] + event_ids if event_ids else [query_embedding]
+
+        results = self.db.execute_sql(sql_query, parameters).fetchall()
+
+        return results
+
+    def update_description(self, event_id: str, description: str) -> None:
+        self.requestor.send_data(
+            EmbeddingsRequestEnum.embed_description.value,
+            {"id": event_id, "description": description},
+        )

frigate/embeddings/embeddings.py

@@ -1,23 +1,26 @@
 """SQLite-vec embeddings database."""
 
 import base64
-import io
 import logging
-import struct
+import os
 import time
-from typing import List, Tuple, Union
 
-from PIL import Image
+from numpy import ndarray
 from playhouse.shortcuts import model_to_dict
 
 from frigate.comms.inter_process import InterProcessRequestor
-from frigate.const import UPDATE_MODEL_STATE
+from frigate.config.semantic_search import SemanticSearchConfig
+from frigate.const import (
+    CONFIG_DIR,
+    UPDATE_EMBEDDINGS_REINDEX_PROGRESS,
+    UPDATE_MODEL_STATE,
+)
 from frigate.db.sqlitevecq import SqliteVecQueueDatabase
 from frigate.models import Event
 from frigate.types import ModelStatusTypesEnum
+from frigate.util.builtin import serialize
 
-from .functions.clip import ClipEmbedding
-from .functions.minilm_l6_v2 import MiniLMEmbedding
+from .functions.onnx import GenericONNXEmbedding, ModelTypeEnum
 
 logger = logging.getLogger(__name__)
 
@@ -53,31 +56,26 @@ def get_metadata(event: Event) -> dict:
     )
 
 
-def serialize(vector: List[float]) -> bytes:
-    """Serializes a list of floats into a compact "raw bytes" format"""
-    return struct.pack("%sf" % len(vector), *vector)
-
-
-def deserialize(bytes_data: bytes) -> List[float]:
-    """Deserializes a compact "raw bytes" format into a list of floats"""
-    return list(struct.unpack("%sf" % (len(bytes_data) // 4), bytes_data))
-
-
 class Embeddings:
     """SQLite-vec embeddings database."""
 
-    def __init__(self, db: SqliteVecQueueDatabase) -> None:
+    def __init__(
+        self, config: SemanticSearchConfig, db: SqliteVecQueueDatabase
+    ) -> None:
+        self.config = config
         self.db = db
         self.requestor = InterProcessRequestor()
 
         # Create tables if they don't exist
-        self._create_tables()
+        self.db.create_embeddings_tables()
 
         models = [
-            "sentence-transformers/all-MiniLM-L6-v2-model.onnx",
-            "sentence-transformers/all-MiniLM-L6-v2-tokenizer",
-            "clip-clip_image_model_vitb32.onnx",
-            "clip-clip_text_model_vitb32.onnx",
+            "jinaai/jina-clip-v1-text_model_fp16.onnx",
+            "jinaai/jina-clip-v1-tokenizer",
+            "jinaai/jina-clip-v1-vision_model_fp16.onnx"
+            if config.model_size == "large"
+            else "jinaai/jina-clip-v1-vision_model_quantized.onnx",
+            "jinaai/jina-clip-v1-preprocessor_config.json",
         ]
 
         for model in models:
@@ -89,171 +87,172 @@ class Embeddings:
                 },
             )
 
-        self.clip_embedding = ClipEmbedding(
-            preferred_providers=["CPUExecutionProvider"]
-        )
-        self.minilm_embedding = MiniLMEmbedding(
-            preferred_providers=["CPUExecutionProvider"],
+        self.text_embedding = GenericONNXEmbedding(
+            model_name="jinaai/jina-clip-v1",
+            model_file="text_model_fp16.onnx",
+            tokenizer_file="tokenizer",
+            download_urls={
+                "text_model_fp16.onnx": "https://huggingface.co/jinaai/jina-clip-v1/resolve/main/onnx/text_model_fp16.onnx",
+            },
+            model_size=config.model_size,
+            model_type=ModelTypeEnum.text,
+            requestor=self.requestor,
+            device="CPU",
         )
 
-    def _create_tables(self):
-        # Create vec0 virtual table for thumbnail embeddings
-        self.db.execute_sql("""
-            CREATE VIRTUAL TABLE IF NOT EXISTS vec_thumbnails USING vec0(
-                id TEXT PRIMARY KEY,
-                thumbnail_embedding FLOAT[512]
-            );
-        """)
-
-        # Create vec0 virtual table for description embeddings
-        self.db.execute_sql("""
-            CREATE VIRTUAL TABLE IF NOT EXISTS vec_descriptions USING vec0(
-                id TEXT PRIMARY KEY,
-                description_embedding FLOAT[384]
-            );
-        """)
-
-    def upsert_thumbnail(self, event_id: str, thumbnail: bytes):
-        # Convert thumbnail bytes to PIL Image
-        image = Image.open(io.BytesIO(thumbnail)).convert("RGB")
-        # Generate embedding using CLIP
-        embedding = self.clip_embedding([image])[0]
-
-        self.db.execute_sql(
-            """
-            INSERT OR REPLACE INTO vec_thumbnails(id, thumbnail_embedding)
-            VALUES(?, ?)
-            """,
-            (event_id, serialize(embedding)),
+        model_file = (
+            "vision_model_fp16.onnx"
+            if self.config.model_size == "large"
+            else "vision_model_quantized.onnx"
         )
 
-        return embedding
-
-    def upsert_description(self, event_id: str, description: str):
-        # Generate embedding using MiniLM
-        embedding = self.minilm_embedding([description])[0]
-
-        self.db.execute_sql(
-            """
-            INSERT OR REPLACE INTO vec_descriptions(id, description_embedding)
-            VALUES(?, ?)
-            """,
-            (event_id, serialize(embedding)),
-        )
-
-        return embedding
-
-    def delete_thumbnail(self, event_ids: List[str]) -> None:
-        ids = ",".join(["?" for _ in event_ids])
-        self.db.execute_sql(
-            f"DELETE FROM vec_thumbnails WHERE id IN ({ids})", event_ids
-        )
-
-    def delete_description(self, event_ids: List[str]) -> None:
-        ids = ",".join(["?" for _ in event_ids])
-        self.db.execute_sql(
-            f"DELETE FROM vec_descriptions WHERE id IN ({ids})", event_ids
-        )
-
-    def search_thumbnail(
-        self, query: Union[Event, str], event_ids: List[str] = None
-    ) -> List[Tuple[str, float]]:
-        if query.__class__ == Event:
-            cursor = self.db.execute_sql(
-                """
-                SELECT thumbnail_embedding FROM vec_thumbnails WHERE id = ?
-                """,
-                [query.id],
-            )
-
-            row = cursor.fetchone() if cursor else None
-
-            if row:
-                query_embedding = deserialize(
-                    row[0]
-                )  # Deserialize the thumbnail embedding
-            else:
-                # If no embedding found, generate it and return it
-                thumbnail = base64.b64decode(query.thumbnail)
-                query_embedding = self.upsert_thumbnail(query.id, thumbnail)
-        else:
-            query_embedding = self.clip_embedding([query])[0]
-
-        sql_query = """
-            SELECT
-                id,
-                distance
-            FROM vec_thumbnails
-            WHERE thumbnail_embedding MATCH ?
-                AND k = 100
-        """
-
-        # Add the IN clause if event_ids is provided and not empty
-        # this is the only filter supported by sqlite-vec as of 0.1.3
-        # but it seems to be broken in this version
-        if event_ids:
-            sql_query += " AND id IN ({})".format(",".join("?" * len(event_ids)))
-
-        # order by distance DESC is not implemented in this version of sqlite-vec
-        # when it's implemented, we can use cosine similarity
-        sql_query += " ORDER BY distance"
-
-        parameters = (
-            [serialize(query_embedding)] + event_ids
-            if event_ids
-            else [serialize(query_embedding)]
-        )
-
-        results = self.db.execute_sql(sql_query, parameters).fetchall()
-
-        return results
-
-    def search_description(
-        self, query_text: str, event_ids: List[str] = None
-    ) -> List[Tuple[str, float]]:
-        query_embedding = self.minilm_embedding([query_text])[0]
-
-        # Prepare the base SQL query
-        sql_query = """
-            SELECT
-                id,
-                distance
-            FROM vec_descriptions
-            WHERE description_embedding MATCH ?
-                AND k = 100
-        """
-
-        # Add the IN clause if event_ids is provided and not empty
-        # this is the only filter supported by sqlite-vec as of 0.1.3
-        # but it seems to be broken in this version
-        if event_ids:
-            sql_query += " AND id IN ({})".format(",".join("?" * len(event_ids)))
-
-        # order by distance DESC is not implemented in this version of sqlite-vec
-        # when it's implemented, we can use cosine similarity
-        sql_query += " ORDER BY distance"
-
-        parameters = (
-            [serialize(query_embedding)] + event_ids
-            if event_ids
-            else [serialize(query_embedding)]
-        )
-
-        results = self.db.execute_sql(sql_query, parameters).fetchall()
-
-        return results
-
-    def reindex(self) -> None:
-        logger.info("Indexing event embeddings...")
-
-        st = time.time()
-        totals = {
-            "thumb": 0,
-            "desc": 0,
+        download_urls = {
+            model_file: f"https://huggingface.co/jinaai/jina-clip-v1/resolve/main/onnx/{model_file}",
+            "preprocessor_config.json": "https://huggingface.co/jinaai/jina-clip-v1/resolve/main/preprocessor_config.json",
         }
 
-        batch_size = 100
+        self.vision_embedding = GenericONNXEmbedding(
+            model_name="jinaai/jina-clip-v1",
+            model_file=model_file,
+            download_urls=download_urls,
+            model_size=config.model_size,
+            model_type=ModelTypeEnum.vision,
+            requestor=self.requestor,
+            device="GPU" if config.model_size == "large" else "CPU",
+        )
+
+    def embed_thumbnail(
+        self, event_id: str, thumbnail: bytes, upsert: bool = True
+    ) -> ndarray:
+        """Embed thumbnail and optionally insert into DB.
+
+        @param: event_id in Events DB
+        @param: thumbnail bytes in jpg format
+        @param: upsert If embedding should be upserted into vec DB
+        """
+        # Convert thumbnail bytes to PIL Image
+        embedding = self.vision_embedding([thumbnail])[0]
+
+        if upsert:
+            self.db.execute_sql(
+                """
+                INSERT OR REPLACE INTO vec_thumbnails(id, thumbnail_embedding)
+                VALUES(?, ?)
+                """,
+                (event_id, serialize(embedding)),
+            )
+
+        return embedding
+
+    def batch_embed_thumbnail(
+        self, event_thumbs: dict[str, bytes], upsert: bool = True
+    ) -> list[ndarray]:
+        """Embed thumbnails and optionally insert into DB.
+
+        @param: event_thumbs Map of Event IDs in DB to thumbnail bytes in jpg format
+        @param: upsert If embedding should be upserted into vec DB
+        """
+        ids = list(event_thumbs.keys())
+        embeddings = self.vision_embedding(list(event_thumbs.values()))
+
+        if upsert:
+            items = []
+
+            for i in range(len(ids)):
+                items.append(ids[i])
+                items.append(serialize(embeddings[i]))
+
+            self.db.execute_sql(
+                """
+                INSERT OR REPLACE INTO vec_thumbnails(id, thumbnail_embedding)
+                VALUES {}
+                """.format(", ".join(["(?, ?)"] * len(ids))),
+                items,
+            )
+
+        return embeddings
+
+    def embed_description(
+        self, event_id: str, description: str, upsert: bool = True
+    ) -> ndarray:
+        embedding = self.text_embedding([description])[0]
+
+        if upsert:
+            self.db.execute_sql(
+                """
+                INSERT OR REPLACE INTO vec_descriptions(id, description_embedding)
+                VALUES(?, ?)
+                """,
+                (event_id, serialize(embedding)),
+            )
+
+        return embedding
+
+    def batch_embed_description(
+        self, event_descriptions: dict[str, str], upsert: bool = True
+    ) -> ndarray:
+        # upsert embeddings one by one to avoid token limit
+        embeddings = []
+
+        for desc in event_descriptions.values():
+            embeddings.append(self.text_embedding([desc])[0])
+
+        if upsert:
+            ids = list(event_descriptions.keys())
+            items = []
+
+            for i in range(len(ids)):
+                items.append(ids[i])
+                items.append(serialize(embeddings[i]))
+
+            self.db.execute_sql(
+                """
+                INSERT OR REPLACE INTO vec_descriptions(id, description_embedding)
+                VALUES {}
+                """.format(", ".join(["(?, ?)"] * len(ids))),
+                items,
+            )
+
+        return embeddings
+
+    def reindex(self) -> None:
+        logger.info("Indexing tracked object embeddings...")
+
+        self.db.drop_embeddings_tables()
+        logger.debug("Dropped embeddings tables.")
+        self.db.create_embeddings_tables()
+        logger.debug("Created embeddings tables.")
+
+        # Delete the saved stats file
+        if os.path.exists(os.path.join(CONFIG_DIR, ".search_stats.json")):
+            os.remove(os.path.join(CONFIG_DIR, ".search_stats.json"))
+
+        st = time.time()
+
+        # Get total count of events to process
+        total_events = (
+            Event.select()
+            .where(
+                (Event.has_clip == True | Event.has_snapshot == True)
+                & Event.thumbnail.is_null(False)
+            )
+            .count()
+        )
+
+        batch_size = 32
         current_page = 1
+
+        totals = {
+            "thumbnails": 0,
+            "descriptions": 0,
+            "processed_objects": total_events - 1 if total_events < batch_size else 0,
+            "total_objects": total_events,
+            "time_remaining": 0 if total_events < batch_size else -1,
+            "status": "indexing",
+        }
+
+        self.requestor.send_data(UPDATE_EMBEDDINGS_REINDEX_PROGRESS, totals)
+
         events = (
             Event.select()
             .where(
@@ -266,14 +265,45 @@ class Embeddings:
 
         while len(events) > 0:
             event: Event
+            batch_thumbs = {}
+            batch_descs = {}
             for event in events:
-                thumbnail = base64.b64decode(event.thumbnail)
-                self.upsert_thumbnail(event.id, thumbnail)
-                totals["thumb"] += 1
-                if description := event.data.get("description", "").strip():
-                    totals["desc"] += 1
-                    self.upsert_description(event.id, description)
+                batch_thumbs[event.id] = base64.b64decode(event.thumbnail)
+                totals["thumbnails"] += 1
 
+                if description := event.data.get("description", "").strip():
+                    batch_descs[event.id] = description
+                    totals["descriptions"] += 1
+
+                totals["processed_objects"] += 1
+
+            # run batch embedding
+            self.batch_embed_thumbnail(batch_thumbs)
+
+            if batch_descs:
+                self.batch_embed_description(batch_descs)
+
+            # report progress every batch so we don't spam the logs
+            progress = (totals["processed_objects"] / total_events) * 100
+            logger.debug(
+                "Processed %d/%d events (%.2f%% complete) | Thumbnails: %d, Descriptions: %d",
+                totals["processed_objects"],
+                total_events,
+                progress,
+                totals["thumbnails"],
+                totals["descriptions"],
+            )
+
+            # Calculate time remaining
+            elapsed_time = time.time() - st
+            avg_time_per_event = elapsed_time / totals["processed_objects"]
+            remaining_events = total_events - totals["processed_objects"]
+            time_remaining = avg_time_per_event * remaining_events
+            totals["time_remaining"] = int(time_remaining)
+
+            self.requestor.send_data(UPDATE_EMBEDDINGS_REINDEX_PROGRESS, totals)
+
+            # Move to the next page
             current_page += 1
             events = (
                 Event.select()
@@ -287,7 +317,10 @@ class Embeddings:
 
         logger.info(
             "Embedded %d thumbnails and %d descriptions in %s seconds",
-            totals["thumb"],
-            totals["desc"],
-            time.time() - st,
+            totals["thumbnails"],
+            totals["descriptions"],
+            round(time.time() - st, 1),
         )
+        totals["status"] = "completed"
+
+        self.requestor.send_data(UPDATE_EMBEDDINGS_REINDEX_PROGRESS, totals)

frigate/embeddings/functions/clip.py

@@ -1,166 +0,0 @@
-import logging
-import os
-from typing import List, Optional, Union
-
-import numpy as np
-import onnxruntime as ort
-from onnx_clip import OnnxClip, Preprocessor, Tokenizer
-from PIL import Image
-
-from frigate.const import MODEL_CACHE_DIR, UPDATE_MODEL_STATE
-from frigate.types import ModelStatusTypesEnum
-from frigate.util.downloader import ModelDownloader
-
-logger = logging.getLogger(__name__)
-
-
-class Clip(OnnxClip):
-    """Override load models to use pre-downloaded models from cache directory."""
-
-    def __init__(
-        self,
-        model: str = "ViT-B/32",
-        batch_size: Optional[int] = None,
-        providers: List[str] = ["CPUExecutionProvider"],
-    ):
-        """
-        Instantiates the model and required encoding classes.
-
-        Args:
-            model: The model to utilize. Currently ViT-B/32 and RN50 are
-                allowed.
-            batch_size: If set, splits the lists in `get_image_embeddings`
-                and `get_text_embeddings` into batches of this size before
-                passing them to the model. The embeddings are then concatenated
-                back together before being returned. This is necessary when
-                passing large amounts of data (perhaps ~100 or more).
-        """
-        allowed_models = ["ViT-B/32", "RN50"]
-        if model not in allowed_models:
-            raise ValueError(f"`model` must be in {allowed_models}. Got {model}.")
-        if model == "ViT-B/32":
-            self.embedding_size = 512
-        elif model == "RN50":
-            self.embedding_size = 1024
-        self.image_model, self.text_model = self._load_models(model, providers)
-        self._tokenizer = Tokenizer()
-        self._preprocessor = Preprocessor()
-        self._batch_size = batch_size
-
-    @staticmethod
-    def _load_models(
-        model: str,
-        providers: List[str],
-    ) -> tuple[ort.InferenceSession, ort.InferenceSession]:
-        """
-        Load models from cache directory.
-        """
-        if model == "ViT-B/32":
-            IMAGE_MODEL_FILE = "clip_image_model_vitb32.onnx"
-            TEXT_MODEL_FILE = "clip_text_model_vitb32.onnx"
-        elif model == "RN50":
-            IMAGE_MODEL_FILE = "clip_image_model_rn50.onnx"
-            TEXT_MODEL_FILE = "clip_text_model_rn50.onnx"
-        else:
-            raise ValueError(f"Unexpected model {model}. No `.onnx` file found.")
-
-        models = []
-        for model_file in [IMAGE_MODEL_FILE, TEXT_MODEL_FILE]:
-            path = os.path.join(MODEL_CACHE_DIR, "clip", model_file)
-            models.append(Clip._load_model(path, providers))
-
-        return models[0], models[1]
-
-    @staticmethod
-    def _load_model(path: str, providers: List[str]):
-        if os.path.exists(path):
-            return ort.InferenceSession(path, providers=providers)
-        else:
-            logger.warning(f"CLIP model file {path} not found.")
-            return None
-
-
-class ClipEmbedding:
-    """Embedding function for CLIP model."""
-
-    def __init__(
-        self,
-        model: str = "ViT-B/32",
-        silent: bool = False,
-        preferred_providers: List[str] = ["CPUExecutionProvider"],
-    ):
-        self.model_name = model
-        self.silent = silent
-        self.preferred_providers = preferred_providers
-        self.model_files = self._get_model_files()
-        self.model = None
-
-        self.downloader = ModelDownloader(
-            model_name="clip",
-            download_path=os.path.join(MODEL_CACHE_DIR, "clip"),
-            file_names=self.model_files,
-            download_func=self._download_model,
-            silent=self.silent,
-        )
-        self.downloader.ensure_model_files()
-
-    def _get_model_files(self):
-        if self.model_name == "ViT-B/32":
-            return ["clip_image_model_vitb32.onnx", "clip_text_model_vitb32.onnx"]
-        elif self.model_name == "RN50":
-            return ["clip_image_model_rn50.onnx", "clip_text_model_rn50.onnx"]
-        else:
-            raise ValueError(
-                f"Unexpected model {self.model_name}. No `.onnx` file found."
-            )
-
-    def _download_model(self, path: str):
-        s3_url = (
-            f"https://lakera-clip.s3.eu-west-1.amazonaws.com/{os.path.basename(path)}"
-        )
-        try:
-            ModelDownloader.download_from_url(s3_url, path, self.silent)
-            self.downloader.requestor.send_data(
-                UPDATE_MODEL_STATE,
-                {
-                    "model": f"{self.model_name}-{os.path.basename(path)}",
-                    "state": ModelStatusTypesEnum.downloaded,
-                },
-            )
-        except Exception:
-            self.downloader.requestor.send_data(
-                UPDATE_MODEL_STATE,
-                {
-                    "model": f"{self.model_name}-{os.path.basename(path)}",
-                    "state": ModelStatusTypesEnum.error,
-                },
-            )
-
-    def _load_model(self):
-        if self.model is None:
-            self.downloader.wait_for_download()
-            self.model = Clip(self.model_name, providers=self.preferred_providers)
-
-    def __call__(self, input: Union[List[str], List[Image.Image]]) -> List[np.ndarray]:
-        self._load_model()
-        if (
-            self.model is None
-            or self.model.image_model is None
-            or self.model.text_model is None
-        ):
-            logger.info(
-                "CLIP model is not fully loaded. Please wait for the download to complete."
-            )
-            return []
-
-        embeddings = []
-        for item in input:
-            if isinstance(item, Image.Image):
-                result = self.model.get_image_embeddings([item])
-                embeddings.append(result[0])
-            elif isinstance(item, str):
-                result = self.model.get_text_embeddings([item])
-                embeddings.append(result[0])
-            else:
-                raise ValueError(f"Unsupported input type: {type(item)}")
-        return embeddings

frigate/embeddings/functions/minilm_l6_v2.py

@@ -1,107 +0,0 @@
-import logging
-import os
-from typing import List
-
-import numpy as np
-import onnxruntime as ort
-
-# importing this without pytorch or others causes a warning
-# https://github.com/huggingface/transformers/issues/27214
-# suppressed by setting env TRANSFORMERS_NO_ADVISORY_WARNINGS=1
-from transformers import AutoTokenizer
-
-from frigate.const import MODEL_CACHE_DIR, UPDATE_MODEL_STATE
-from frigate.types import ModelStatusTypesEnum
-from frigate.util.downloader import ModelDownloader
-
-logger = logging.getLogger(__name__)
-
-
-class MiniLMEmbedding:
-    """Embedding function for ONNX MiniLM-L6 model."""
-
-    DOWNLOAD_PATH = f"{MODEL_CACHE_DIR}/all-MiniLM-L6-v2"
-    MODEL_NAME = "sentence-transformers/all-MiniLM-L6-v2"
-    IMAGE_MODEL_FILE = "model.onnx"
-    TOKENIZER_FILE = "tokenizer"
-
-    def __init__(self, preferred_providers=["CPUExecutionProvider"]):
-        self.preferred_providers = preferred_providers
-        self.tokenizer = None
-        self.session = None
-
-        self.downloader = ModelDownloader(
-            model_name=self.MODEL_NAME,
-            download_path=self.DOWNLOAD_PATH,
-            file_names=[self.IMAGE_MODEL_FILE, self.TOKENIZER_FILE],
-            download_func=self._download_model,
-        )
-        self.downloader.ensure_model_files()
-
-    def _download_model(self, path: str):
-        try:
-            if os.path.basename(path) == self.IMAGE_MODEL_FILE:
-                s3_url = f"https://huggingface.co/sentence-transformers/all-MiniLM-L6-v2/resolve/main/onnx/{self.IMAGE_MODEL_FILE}"
-                ModelDownloader.download_from_url(s3_url, path)
-            elif os.path.basename(path) == self.TOKENIZER_FILE:
-                logger.info("Downloading MiniLM tokenizer")
-                tokenizer = AutoTokenizer.from_pretrained(
-                    self.MODEL_NAME, clean_up_tokenization_spaces=True
-                )
-                tokenizer.save_pretrained(path)
-
-            self.downloader.requestor.send_data(
-                UPDATE_MODEL_STATE,
-                {
-                    "model": f"{self.MODEL_NAME}-{os.path.basename(path)}",
-                    "state": ModelStatusTypesEnum.downloaded,
-                },
-            )
-        except Exception:
-            self.downloader.requestor.send_data(
-                UPDATE_MODEL_STATE,
-                {
-                    "model": f"{self.MODEL_NAME}-{os.path.basename(path)}",
-                    "state": ModelStatusTypesEnum.error,
-                },
-            )
-
-    def _load_model_and_tokenizer(self):
-        if self.tokenizer is None or self.session is None:
-            self.downloader.wait_for_download()
-            self.tokenizer = self._load_tokenizer()
-            self.session = self._load_model(
-                os.path.join(self.DOWNLOAD_PATH, self.IMAGE_MODEL_FILE),
-                self.preferred_providers,
-            )
-
-    def _load_tokenizer(self):
-        tokenizer_path = os.path.join(self.DOWNLOAD_PATH, self.TOKENIZER_FILE)
-        return AutoTokenizer.from_pretrained(
-            tokenizer_path, clean_up_tokenization_spaces=True
-        )
-
-    def _load_model(self, path: str, providers: List[str]):
-        if os.path.exists(path):
-            return ort.InferenceSession(path, providers=providers)
-        else:
-            logger.warning(f"MiniLM model file {path} not found.")
-            return None
-
-    def __call__(self, texts: List[str]) -> List[np.ndarray]:
-        self._load_model_and_tokenizer()
-
-        if self.session is None or self.tokenizer is None:
-            logger.error("MiniLM model or tokenizer is not loaded.")
-            return []
-
-        inputs = self.tokenizer(
-            texts, padding=True, truncation=True, return_tensors="np"
-        )
-        input_names = [input.name for input in self.session.get_inputs()]
-        onnx_inputs = {name: inputs[name] for name in input_names if name in inputs}
-
-        outputs = self.session.run(None, onnx_inputs)
-        embeddings = outputs[0].mean(axis=1)
-
-        return [embedding for embedding in embeddings]

frigate/embeddings/functions/onnx.py

@@ -0,0 +1,216 @@
+import logging
+import os
+import warnings
+from enum import Enum
+from io import BytesIO
+from typing import Dict, List, Optional, Union
+
+import numpy as np
+import requests
+from PIL import Image
+
+# importing this without pytorch or others causes a warning
+# https://github.com/huggingface/transformers/issues/27214
+# suppressed by setting env TRANSFORMERS_NO_ADVISORY_WARNINGS=1
+from transformers import AutoFeatureExtractor, AutoTokenizer
+from transformers.utils.logging import disable_progress_bar
+
+from frigate.comms.inter_process import InterProcessRequestor
+from frigate.const import MODEL_CACHE_DIR, UPDATE_MODEL_STATE
+from frigate.types import ModelStatusTypesEnum
+from frigate.util.downloader import ModelDownloader
+from frigate.util.model import ONNXModelRunner
+
+warnings.filterwarnings(
+    "ignore",
+    category=FutureWarning,
+    message="The class CLIPFeatureExtractor is deprecated",
+)
+
+# disables the progress bar for downloading tokenizers and feature extractors
+disable_progress_bar()
+logger = logging.getLogger(__name__)
+
+
+class ModelTypeEnum(str, Enum):
+    face = "face"
+    vision = "vision"
+    text = "text"
+
+
+class GenericONNXEmbedding:
+    """Generic embedding function for ONNX models (text and vision)."""
+
+    def __init__(
+        self,
+        model_name: str,
+        model_file: str,
+        download_urls: Dict[str, str],
+        model_size: str,
+        model_type: str,
+        requestor: InterProcessRequestor,
+        tokenizer_file: Optional[str] = None,
+        device: str = "AUTO",
+    ):
+        self.model_name = model_name
+        self.model_file = model_file
+        self.tokenizer_file = tokenizer_file
+        self.requestor = requestor
+        self.download_urls = download_urls
+        self.model_type = model_type  # 'text' or 'vision'
+        self.model_size = model_size
+        self.device = device
+        self.download_path = os.path.join(MODEL_CACHE_DIR, self.model_name)
+        self.tokenizer = None
+        self.feature_extractor = None
+        self.runner = None
+        files_names = list(self.download_urls.keys()) + (
+            [self.tokenizer_file] if self.tokenizer_file else []
+        )
+
+        if not all(
+            os.path.exists(os.path.join(self.download_path, n)) for n in files_names
+        ):
+            logger.debug(f"starting model download for {self.model_name}")
+            self.downloader = ModelDownloader(
+                model_name=self.model_name,
+                download_path=self.download_path,
+                file_names=files_names,
+                download_func=self._download_model,
+            )
+            self.downloader.ensure_model_files()
+        else:
+            self.downloader = None
+            ModelDownloader.mark_files_state(
+                self.requestor,
+                self.model_name,
+                files_names,
+                ModelStatusTypesEnum.downloaded,
+            )
+            self._load_model_and_tokenizer()
+            logger.debug(f"models are already downloaded for {self.model_name}")
+
+    def _download_model(self, path: str):
+        try:
+            file_name = os.path.basename(path)
+            if file_name in self.download_urls:
+                ModelDownloader.download_from_url(self.download_urls[file_name], path)
+            elif (
+                file_name == self.tokenizer_file
+                and self.model_type == ModelTypeEnum.text
+            ):
+                if not os.path.exists(path + "/" + self.model_name):
+                    logger.info(f"Downloading {self.model_name} tokenizer")
+                tokenizer = AutoTokenizer.from_pretrained(
+                    self.model_name,
+                    trust_remote_code=True,
+                    cache_dir=f"{MODEL_CACHE_DIR}/{self.model_name}/tokenizer",
+                    clean_up_tokenization_spaces=True,
+                )
+                tokenizer.save_pretrained(path)
+
+            self.downloader.requestor.send_data(
+                UPDATE_MODEL_STATE,
+                {
+                    "model": f"{self.model_name}-{file_name}",
+                    "state": ModelStatusTypesEnum.downloaded,
+                },
+            )
+        except Exception:
+            self.downloader.requestor.send_data(
+                UPDATE_MODEL_STATE,
+                {
+                    "model": f"{self.model_name}-{file_name}",
+                    "state": ModelStatusTypesEnum.error,
+                },
+            )
+
+    def _load_model_and_tokenizer(self):
+        if self.runner is None:
+            if self.downloader:
+                self.downloader.wait_for_download()
+            if self.model_type == ModelTypeEnum.text:
+                self.tokenizer = self._load_tokenizer()
+            else:
+                self.feature_extractor = self._load_feature_extractor()
+            self.runner = ONNXModelRunner(
+                os.path.join(self.download_path, self.model_file),
+                self.device,
+                self.model_size,
+            )
+
+    def _load_tokenizer(self):
+        tokenizer_path = os.path.join(f"{MODEL_CACHE_DIR}/{self.model_name}/tokenizer")
+        return AutoTokenizer.from_pretrained(
+            self.model_name,
+            cache_dir=tokenizer_path,
+            trust_remote_code=True,
+            clean_up_tokenization_spaces=True,
+        )
+
+    def _load_feature_extractor(self):
+        return AutoFeatureExtractor.from_pretrained(
+            f"{MODEL_CACHE_DIR}/{self.model_name}",
+        )
+
+    def _preprocess_inputs(self, raw_inputs: any) -> any:
+        if self.model_type == ModelTypeEnum.text:
+            max_length = max(len(self.tokenizer.encode(text)) for text in raw_inputs)
+            return [
+                self.tokenizer(
+                    text,
+                    padding="max_length",
+                    truncation=True,
+                    max_length=max_length,
+                    return_tensors="np",
+                )
+                for text in raw_inputs
+            ]
+        elif self.model_type == ModelTypeEnum.vision:
+            processed_images = [self._process_image(img) for img in raw_inputs]
+            return [
+                self.feature_extractor(images=image, return_tensors="np")
+                for image in processed_images
+            ]
+        else:
+            raise ValueError(f"Unable to preprocess inputs for {self.model_type}")
+
+    def _process_image(self, image):
+        if isinstance(image, str):
+            if image.startswith("http"):
+                response = requests.get(image)
+                image = Image.open(BytesIO(response.content)).convert("RGB")
+        elif isinstance(image, bytes):
+            image = Image.open(BytesIO(image)).convert("RGB")
+
+        return image
+
+    def __call__(
+        self, inputs: Union[List[str], List[Image.Image], List[str]]
+    ) -> List[np.ndarray]:
+        self._load_model_and_tokenizer()
+        if self.runner is None or (
+            self.tokenizer is None and self.feature_extractor is None
+        ):
+            logger.error(
+                f"{self.model_name} model or tokenizer/feature extractor is not loaded."
+            )
+            return []
+
+        processed_inputs = self._preprocess_inputs(inputs)
+        input_names = self.runner.get_input_names()
+        onnx_inputs = {name: [] for name in input_names}
+        input: dict[str, any]
+        for input in processed_inputs:
+            for key, value in input.items():
+                if key in input_names:
+                    onnx_inputs[key].append(value[0])
+
+        for key in input_names:
+            if onnx_inputs.get(key):
+                onnx_inputs[key] = np.stack(onnx_inputs[key])
+            else:
+                logger.warning(f"Expected input '{key}' not found in onnx_inputs")
+
+        embeddings = self.runner.run(onnx_inputs)[0]
+        return [embedding for embedding in embeddings]

frigate/embeddings/maintainer.py

@@ -12,6 +12,7 @@ import numpy as np
 from peewee import DoesNotExist
 from playhouse.sqliteq import SqliteQueueDatabase
 
+from frigate.comms.embeddings_updater import EmbeddingsRequestEnum, EmbeddingsResponder
 from frigate.comms.event_metadata_updater import (
     EventMetadataSubscriber,
     EventMetadataTypeEnum,
@@ -23,12 +24,16 @@ from frigate.const import CLIPS_DIR, UPDATE_EVENT_DESCRIPTION
 from frigate.events.types import EventTypeEnum
 from frigate.genai import get_genai_client
 from frigate.models import Event
+from frigate.types import TrackedObjectUpdateTypesEnum
+from frigate.util.builtin import serialize
 from frigate.util.image import SharedMemoryFrameManager, calculate_region
 
 from .embeddings import Embeddings
 
 logger = logging.getLogger(__name__)
 
+MAX_THUMBNAILS = 10
+
 
 class EmbeddingMaintainer(threading.Thread):
     """Handle embedding queue and post event updates."""
@@ -39,25 +44,31 @@ class EmbeddingMaintainer(threading.Thread):
         config: FrigateConfig,
         stop_event: MpEvent,
     ) -> None:
-        threading.Thread.__init__(self)
-        self.name = "embeddings_maintainer"
+        super().__init__(name="embeddings_maintainer")
         self.config = config
-        self.embeddings = Embeddings(db)
+        self.embeddings = Embeddings(config.semantic_search, db)
+
+        # Check if we need to re-index events
+        if config.semantic_search.reindex:
+            self.embeddings.reindex()
+
         self.event_subscriber = EventUpdateSubscriber()
         self.event_end_subscriber = EventEndSubscriber()
         self.event_metadata_subscriber = EventMetadataSubscriber(
             EventMetadataTypeEnum.regenerate_description
         )
+        self.embeddings_responder = EmbeddingsResponder()
         self.frame_manager = SharedMemoryFrameManager()
         # create communication for updating event descriptions
         self.requestor = InterProcessRequestor()
         self.stop_event = stop_event
         self.tracked_events = {}
-        self.genai_client = get_genai_client(config.genai)
+        self.genai_client = get_genai_client(config)
 
     def run(self) -> None:
         """Maintain a SQLite-vec database for semantic search."""
         while not self.stop_event.is_set():
+            self._process_requests()
             self._process_updates()
             self._process_finalized()
             self._process_event_metadata()
@@ -65,41 +76,86 @@ class EmbeddingMaintainer(threading.Thread):
         self.event_subscriber.stop()
         self.event_end_subscriber.stop()
         self.event_metadata_subscriber.stop()
+        self.embeddings_responder.stop()
         self.requestor.stop()
         logger.info("Exiting embeddings maintenance...")
 
+    def _process_requests(self) -> None:
+        """Process embeddings requests"""
+
+        def _handle_request(topic: str, data: str) -> str:
+            try:
+                if topic == EmbeddingsRequestEnum.embed_description.value:
+                    return serialize(
+                        self.embeddings.embed_description(
+                            data["id"], data["description"]
+                        ),
+                        pack=False,
+                    )
+                elif topic == EmbeddingsRequestEnum.embed_thumbnail.value:
+                    thumbnail = base64.b64decode(data["thumbnail"])
+                    return serialize(
+                        self.embeddings.embed_thumbnail(data["id"], thumbnail),
+                        pack=False,
+                    )
+                elif topic == EmbeddingsRequestEnum.generate_search.value:
+                    return serialize(
+                        self.embeddings.text_embedding([data])[0], pack=False
+                    )
+            except Exception as e:
+                logger.error(f"Unable to handle embeddings request {e}")
+
+        self.embeddings_responder.check_for_request(_handle_request)
+
     def _process_updates(self) -> None:
         """Process event updates"""
-        update = self.event_subscriber.check_for_update()
+        update = self.event_subscriber.check_for_update(timeout=0.1)
 
         if update is None:
             return
 
-        source_type, _, camera, data = update
+        source_type, _, camera, frame_name, data = update
 
         if not camera or source_type != EventTypeEnum.tracked_object:
             return
 
         camera_config = self.config.cameras[camera]
+        # no need to save our own thumbnails if genai is not enabled
+        # or if the object has become stationary
+        if (
+            not camera_config.genai.enabled
+            or self.genai_client is None
+            or data["stationary"]
+        ):
+            return
+
         if data["id"] not in self.tracked_events:
             self.tracked_events[data["id"]] = []
 
         # Create our own thumbnail based on the bounding box and the frame time
         try:
-            frame_id = f"{camera}{data['frame_time']}"
-            yuv_frame = self.frame_manager.get(frame_id, camera_config.frame_shape_yuv)
+            yuv_frame = self.frame_manager.get(
+                frame_name, camera_config.frame_shape_yuv
+            )
 
             if yuv_frame is not None:
                 data["thumbnail"] = self._create_thumbnail(yuv_frame, data["box"])
+
+                # Limit the number of thumbnails saved
+                if len(self.tracked_events[data["id"]]) >= MAX_THUMBNAILS:
+                    # Always keep the first thumbnail for the event
+                    self.tracked_events[data["id"]].pop(1)
+
                 self.tracked_events[data["id"]].append(data)
-                self.frame_manager.close(frame_id)
+
+                self.frame_manager.close(frame_name)
         except FileNotFoundError:
             pass
 
     def _process_finalized(self) -> None:
         """Process the end of an event."""
         while True:
-            ended = self.event_end_subscriber.check_for_update()
+            ended = self.event_end_subscriber.check_for_update(timeout=0.1)
 
             if ended == None:
                 break
@@ -136,9 +192,6 @@ class EmbeddingMaintainer(threading.Thread):
                         or set(event.zones) & set(camera_config.genai.required_zones)
                     )
                 ):
-                    logger.debug(
-                        f"Description generation for {event}, has_snapshot: {event.has_snapshot}"
-                    )
                     if event.has_snapshot and camera_config.genai.use_snapshot:
                         with open(
                             os.path.join(CLIPS_DIR, f"{event.camera}-{event.id}.jpg"),
@@ -192,7 +245,7 @@ class EmbeddingMaintainer(threading.Thread):
     def _process_event_metadata(self):
         # Check for regenerate description requests
         (topic, event_id, source) = self.event_metadata_subscriber.check_for_update(
-            timeout=1
+            timeout=0.1
         )
 
         if topic is None:
@@ -219,14 +272,14 @@ class EmbeddingMaintainer(threading.Thread):
 
     def _embed_thumbnail(self, event_id: str, thumbnail: bytes) -> None:
         """Embed the thumbnail for an event."""
-        self.embeddings.upsert_thumbnail(event_id, thumbnail)
+        self.embeddings.embed_thumbnail(event_id, thumbnail)
 
     def _embed_description(self, event: Event, thumbnails: list[bytes]) -> None:
         """Embed the description for an event."""
         camera_config = self.config.cameras[event.camera]
 
         description = self.genai_client.generate_description(
-            camera_config, thumbnails, event.label
+            camera_config, thumbnails, event
         )
 
         if not description:
@@ -236,11 +289,15 @@ class EmbeddingMaintainer(threading.Thread):
         # fire and forget description update
         self.requestor.send_data(
             UPDATE_EVENT_DESCRIPTION,
-            {"id": event.id, "description": description},
+            {
+                "type": TrackedObjectUpdateTypesEnum.description,
+                "id": event.id,
+                "description": description,
+            },
         )
 
-        # Encode the description
-        self.embeddings.upsert_description(event.id, description)
+        # Embed the description
+        self.embeddings.embed_description(event.id, description)
 
         logger.debug(
             "Generated description for %s (%d images): %s",

frigate/embeddings/util.py

@@ -20,10 +20,11 @@ class ZScoreNormalization:
 
     @property
     def stddev(self):
-        return math.sqrt(self.variance)
+        return math.sqrt(self.variance) if self.variance > 0 else 0.0
 
-    def normalize(self, distances: list[float]):
-        self._update(distances)
+    def normalize(self, distances: list[float], save_stats: bool):
+        if save_stats:
+            self._update(distances)
         if self.stddev == 0:
             return distances
         return [

frigate/events/audio.py

@@ -64,6 +64,8 @@ def get_ffmpeg_command(ffmpeg: FfmpegConfig) -> list[str]:
 
 
 class AudioProcessor(util.Process):
+    name = "frigate.audio_manager"
+
     def __init__(
         self,
         cameras: list[CameraConfig],

frigate/events/cleanup.py

@@ -8,11 +8,9 @@ from enum import Enum
 from multiprocessing.synchronize import Event as MpEvent
 from pathlib import Path
 
-from playhouse.sqliteq import SqliteQueueDatabase
-
 from frigate.config import FrigateConfig
 from frigate.const import CLIPS_DIR
-from frigate.embeddings.embeddings import Embeddings
+from frigate.db.sqlitevecq import SqliteVecQueueDatabase
 from frigate.models import Event, Timeline
 
 logger = logging.getLogger(__name__)
@@ -23,9 +21,12 @@ class EventCleanupType(str, Enum):
     snapshots = "snapshots"
 
 
+CHUNK_SIZE = 50
+
+
 class EventCleanup(threading.Thread):
     def __init__(
-        self, config: FrigateConfig, stop_event: MpEvent, db: SqliteQueueDatabase
+        self, config: FrigateConfig, stop_event: MpEvent, db: SqliteVecQueueDatabase
     ):
         super().__init__(name="event_cleanup")
         self.config = config
@@ -35,9 +36,6 @@ class EventCleanup(threading.Thread):
         self.removed_camera_labels: list[str] = None
         self.camera_labels: dict[str, dict[str, any]] = {}
 
-        if self.config.semantic_search.enabled:
-            self.embeddings = Embeddings(self.db)
-
     def get_removed_camera_labels(self) -> list[Event]:
         """Get a list of distinct labels for removed cameras."""
         if self.removed_camera_labels is None:
@@ -112,6 +110,7 @@ class EventCleanup(threading.Thread):
                 .namedtuples()
                 .iterator()
             )
+            logger.debug(f"{len(expired_events)} events can be expired")
             # delete the media from disk
             for expired in expired_events:
                 media_name = f"{expired.camera}-{expired.id}"
@@ -130,13 +129,34 @@ class EventCleanup(threading.Thread):
                     logger.warning(f"Unable to delete event images: {e}")
 
             # update the clips attribute for the db entry
-            update_query = Event.update(update_params).where(
+            query = Event.select(Event.id).where(
                 Event.camera.not_in(self.camera_keys),
                 Event.start_time < expire_after,
                 Event.label == event.label,
                 Event.retain_indefinitely == False,
             )
-            update_query.execute()
+
+            events_to_update = []
+
+            for batch in query.iterator():
+                events_to_update.extend([event.id for event in batch])
+                if len(events_to_update) >= CHUNK_SIZE:
+                    logger.debug(
+                        f"Updating {update_params} for {len(events_to_update)} events"
+                    )
+                    Event.update(update_params).where(
+                        Event.id << events_to_update
+                    ).execute()
+                    events_to_update = []
+
+            # Update any remaining events
+            if events_to_update:
+                logger.debug(
+                    f"Updating clips/snapshots attribute for {len(events_to_update)} events"
+                )
+                Event.update(update_params).where(
+                    Event.id << events_to_update
+                ).execute()
 
         events_to_update = []
 
@@ -201,7 +221,11 @@ class EventCleanup(threading.Thread):
                             logger.warning(f"Unable to delete event images: {e}")
 
         # update the clips attribute for the db entry
-        Event.update(update_params).where(Event.id << events_to_update).execute()
+        for i in range(0, len(events_to_update), CHUNK_SIZE):
+            batch = events_to_update[i : i + CHUNK_SIZE]
+            logger.debug(f"Updating {update_params} for {len(batch)} events")
+            Event.update(update_params).where(Event.id << batch).execute()
+
         return events_to_update
 
     def run(self) -> None:
@@ -227,15 +251,16 @@ class EventCleanup(threading.Thread):
                 .iterator()
             )
             events_to_delete = [e.id for e in events]
+            logger.debug(f"Found {len(events_to_delete)} events that can be expired")
             if len(events_to_delete) > 0:
-                chunk_size = 50
-                for i in range(0, len(events_to_delete), chunk_size):
-                    chunk = events_to_delete[i : i + chunk_size]
+                for i in range(0, len(events_to_delete), CHUNK_SIZE):
+                    chunk = events_to_delete[i : i + CHUNK_SIZE]
+                    logger.debug(f"Deleting {len(chunk)} events from the database")
                     Event.delete().where(Event.id << chunk).execute()
 
                     if self.config.semantic_search.enabled:
-                        self.embeddings.delete_description(chunk)
-                        self.embeddings.delete_thumbnail(chunk)
+                        self.db.delete_embeddings_description(event_ids=chunk)
+                        self.db.delete_embeddings_thumbnail(event_ids=chunk)
                         logger.debug(f"Deleted {len(events_to_delete)} embeddings")
 
         logger.info("Exiting event cleanup...")

frigate/events/external.py

@@ -70,7 +70,7 @@ class ExternalEventProcessor:
                     "sub_label": sub_label,
                     "score": score,
                     "camera": camera,
-                    "start_time": now,
+                    "start_time": now - camera_config.record.event_pre_capture,
                     "end_time": end,
                     "thumbnail": thumbnail,
                     "has_clip": camera_config.record.enabled and include_recording,

frigate/events/maintainer.py

@@ -75,7 +75,7 @@ class EventProcessor(threading.Thread):
             if update == None:
                 continue
 
-            source_type, event_type, camera, event_data = update
+            source_type, event_type, camera, _, event_data = update
 
             logger.debug(
                 f"Event received: {source_type} {event_type} {camera} {event_data['id']}"

frigate/genai/__init__.py

@@ -1,10 +1,16 @@
 """Generative AI module for Frigate."""
 
 import importlib
+import logging
 import os
 from typing import Optional
 
-from frigate.config import CameraConfig, GenAIConfig, GenAIProviderEnum
+from playhouse.shortcuts import model_to_dict
+
+from frigate.config import CameraConfig, FrigateConfig, GenAIConfig, GenAIProviderEnum
+from frigate.models import Event
+
+logger = logging.getLogger(__name__)
 
 PROVIDERS = {}
 
@@ -31,12 +37,14 @@ class GenAIClient:
         self,
         camera_config: CameraConfig,
         thumbnails: list[bytes],
-        label: str,
+        event: Event,
     ) -> Optional[str]:
         """Generate a description for the frame."""
         prompt = camera_config.genai.object_prompts.get(
-            label, camera_config.genai.prompt
-        )
+            event.label,
+            camera_config.genai.prompt,
+        ).format(**model_to_dict(event))
+        logger.debug(f"Sending images to genai provider with prompt: {prompt}")
         return self._send(prompt, thumbnails)
 
     def _init_provider(self):
@@ -48,13 +56,19 @@ class GenAIClient:
         return None
 
 
-def get_genai_client(genai_config: GenAIConfig) -> Optional[GenAIClient]:
+def get_genai_client(config: FrigateConfig) -> Optional[GenAIClient]:
     """Get the GenAI client."""
-    if genai_config.enabled:
+    genai_config = config.genai
+    genai_cameras = [
+        c for c in config.cameras.values() if c.enabled and c.genai.enabled
+    ]
+
+    if genai_cameras:
         load_providers()
         provider = PROVIDERS.get(genai_config.provider)
         if provider:
             return provider(genai_config)
+
     return None
 
 

frigate/genai/ollama.py

@@ -21,12 +21,20 @@ class OllamaClient(GenAIClient):
 
     def _init_provider(self):
         """Initialize the client."""
-        client = ApiClient(host=self.genai_config.base_url, timeout=self.timeout)
-        response = client.pull(self.genai_config.model)
-        if response["status"] != "success":
-            logger.error("Failed to pull %s model from Ollama", self.genai_config.model)
+        try:
+            client = ApiClient(host=self.genai_config.base_url, timeout=self.timeout)
+            # ensure the model is available locally
+            response = client.show(self.genai_config.model)
+            if response.get("error"):
+                logger.error(
+                    "Ollama error: %s",
+                    response["error"],
+                )
+                return None
+            return client
+        except Exception as e:
+            logger.warning("Error initializing Ollama: %s", str(e))
             return None
-        return client
 
     def _send(self, prompt: str, images: list[bytes]) -> Optional[str]:
         """Submit a request to Ollama"""

frigate/models.py

@@ -93,7 +93,7 @@ class ReviewSegment(Model):  # type: ignore[misc]
     start_time = DateTimeField()
     end_time = DateTimeField()
     has_been_reviewed = BooleanField(default=False)
-    severity = CharField(max_length=30)  # alert, detection, significant_motion
+    severity = CharField(max_length=30)  # alert, detection
     thumb_path = CharField(unique=True)
     data = JSONField()  # additional data about detection like list of labels, zone, areas of significant motion
 

frigate/mypy.ini

@@ -59,3 +59,7 @@ ignore_errors = false
 [mypy-frigate.watchdog]
 ignore_errors = false
 disallow_untyped_calls = false
+
+
+[mypy-frigate.service_manager.*]
+ignore_errors = false

frigate/object_detection.py

@@ -12,10 +12,14 @@ from setproctitle import setproctitle
 
 import frigate.util as util
 from frigate.detectors import create_detector
-from frigate.detectors.detector_config import BaseDetectorConfig, InputTensorEnum
+from frigate.detectors.detector_config import (
+    BaseDetectorConfig,
+    InputDTypeEnum,
+    InputTensorEnum,
+)
 from frigate.detectors.plugins.rocm import DETECTOR_KEY as ROCM_DETECTOR_KEY
 from frigate.util.builtin import EventsPerSecond, load_labels
-from frigate.util.image import SharedMemoryFrameManager
+from frigate.util.image import SharedMemoryFrameManager, UntrackedSharedMemory
 from frigate.util.services import listen
 
 logger = logging.getLogger(__name__)
@@ -55,12 +59,15 @@ class LocalObjectDetector(ObjectDetector):
                 self.input_transform = tensor_transform(
                     detector_config.model.input_tensor
                 )
+
+            self.dtype = detector_config.model.input_dtype
         else:
             self.input_transform = None
+            self.dtype = InputDTypeEnum.int
 
         self.detect_api = create_detector(detector_config)
 
-    def detect(self, tensor_input, threshold=0.4):
+    def detect(self, tensor_input: np.ndarray, threshold=0.4):
         detections = []
 
         raw_detections = self.detect_raw(tensor_input)
@@ -77,9 +84,14 @@ class LocalObjectDetector(ObjectDetector):
         self.fps.update()
         return detections
 
-    def detect_raw(self, tensor_input):
+    def detect_raw(self, tensor_input: np.ndarray):
         if self.input_transform:
             tensor_input = np.transpose(tensor_input, self.input_transform)
+
+        if self.dtype == InputDTypeEnum.float:
+            tensor_input = tensor_input.astype(np.float32)
+            tensor_input /= 255
+
         return self.detect_api.detect_raw(tensor_input=tensor_input)
 
 
@@ -110,7 +122,7 @@ def run_detector(
 
     outputs = {}
     for name in out_events.keys():
-        out_shm = mp.shared_memory.SharedMemory(name=f"out-{name}", create=False)
+        out_shm = UntrackedSharedMemory(name=f"out-{name}", create=False)
         out_np = np.ndarray((20, 6), dtype=np.float32, buffer=out_shm.buf)
         outputs[name] = {"shm": out_shm, "np": out_np}
 
@@ -200,15 +212,13 @@ class RemoteObjectDetector:
         self.detection_queue = detection_queue
         self.event = event
         self.stop_event = stop_event
-        self.shm = mp.shared_memory.SharedMemory(name=self.name, create=False)
+        self.shm = UntrackedSharedMemory(name=self.name, create=False)
         self.np_shm = np.ndarray(
             (1, model_config.height, model_config.width, 3),
             dtype=np.uint8,
             buffer=self.shm.buf,
         )
-        self.out_shm = mp.shared_memory.SharedMemory(
-            name=f"out-{self.name}", create=False
-        )
+        self.out_shm = UntrackedSharedMemory(name=f"out-{self.name}", create=False)
         self.out_np_shm = np.ndarray((20, 6), dtype=np.float32, buffer=self.out_shm.buf)
 
     def detect(self, tensor_input, threshold=0.4):

frigate/object_processing.py

@@ -1,4 +1,3 @@
-import base64
 import datetime
 import json
 import logging
@@ -7,7 +6,6 @@ import queue
 import threading
 from collections import Counter, defaultdict
 from multiprocessing.synchronize import Event as MpEvent
-from statistics import median
 from typing import Callable
 
 import cv2
@@ -18,7 +16,6 @@ from frigate.comms.dispatcher import Dispatcher
 from frigate.comms.events_updater import EventEndSubscriber, EventUpdatePublisher
 from frigate.comms.inter_process import InterProcessRequestor
 from frigate.config import (
-    CameraConfig,
     FrigateConfig,
     MqttConfig,
     RecordConfig,
@@ -28,458 +25,18 @@ from frigate.config import (
 from frigate.const import CLIPS_DIR, UPDATE_CAMERA_ACTIVITY
 from frigate.events.types import EventStateEnum, EventTypeEnum
 from frigate.ptz.autotrack import PtzAutoTrackerThread
+from frigate.track.tracked_object import TrackedObject
 from frigate.util.image import (
     SharedMemoryFrameManager,
-    area,
-    calculate_region,
     draw_box_with_label,
     draw_timestamp,
+    is_better_thumbnail,
     is_label_printable,
 )
 
 logger = logging.getLogger(__name__)
 
 
-def on_edge(box, frame_shape):
-    if (
-        box[0] == 0
-        or box[1] == 0
-        or box[2] == frame_shape[1] - 1
-        or box[3] == frame_shape[0] - 1
-    ):
-        return True
-
-
-def has_better_attr(current_thumb, new_obj, attr_label) -> bool:
-    max_new_attr = max(
-        [0]
-        + [area(a["box"]) for a in new_obj["attributes"] if a["label"] == attr_label]
-    )
-    max_current_attr = max(
-        [0]
-        + [
-            area(a["box"])
-            for a in current_thumb["attributes"]
-            if a["label"] == attr_label
-        ]
-    )
-
-    # if the thumb has a higher scoring attr
-    return max_new_attr > max_current_attr
-
-
-def is_better_thumbnail(label, current_thumb, new_obj, frame_shape) -> bool:
-    # larger is better
-    # cutoff images are less ideal, but they should also be smaller?
-    # better scores are obviously better too
-
-    # check face on person
-    if label == "person":
-        if has_better_attr(current_thumb, new_obj, "face"):
-            return True
-        # if the current thumb has a face attr, dont update unless it gets better
-        if any([a["label"] == "face" for a in current_thumb["attributes"]]):
-            return False
-
-    # check license_plate on car
-    if label == "car":
-        if has_better_attr(current_thumb, new_obj, "license_plate"):
-            return True
-        # if the current thumb has a license_plate attr, dont update unless it gets better
-        if any([a["label"] == "license_plate" for a in current_thumb["attributes"]]):
-            return False
-
-    # if the new_thumb is on an edge, and the current thumb is not
-    if on_edge(new_obj["box"], frame_shape) and not on_edge(
-        current_thumb["box"], frame_shape
-    ):
-        return False
-
-    # if the score is better by more than 5%
-    if new_obj["score"] > current_thumb["score"] + 0.05:
-        return True
-
-    # if the area is 10% larger
-    if new_obj["area"] > current_thumb["area"] * 1.1:
-        return True
-
-    return False
-
-
-class TrackedObject:
-    def __init__(
-        self,
-        camera,
-        colormap,
-        camera_config: CameraConfig,
-        frame_cache,
-        obj_data: dict[str, any],
-    ):
-        # set the score history then remove as it is not part of object state
-        self.score_history = obj_data["score_history"]
-        del obj_data["score_history"]
-
-        self.obj_data = obj_data
-        self.camera = camera
-        self.colormap = colormap
-        self.camera_config = camera_config
-        self.frame_cache = frame_cache
-        self.zone_presence: dict[str, int] = {}
-        self.zone_loitering: dict[str, int] = {}
-        self.current_zones = []
-        self.entered_zones = []
-        self.attributes = defaultdict(float)
-        self.false_positive = True
-        self.has_clip = False
-        self.has_snapshot = False
-        self.top_score = self.computed_score = 0.0
-        self.thumbnail_data = None
-        self.last_updated = 0
-        self.last_published = 0
-        self.frame = None
-        self.active = True
-        self.previous = self.to_dict()
-
-    def _is_false_positive(self):
-        # once a true positive, always a true positive
-        if not self.false_positive:
-            return False
-
-        threshold = self.camera_config.objects.filters[self.obj_data["label"]].threshold
-        return self.computed_score < threshold
-
-    def compute_score(self):
-        """get median of scores for object."""
-        return median(self.score_history)
-
-    def update(self, current_frame_time: float, obj_data, has_valid_frame: bool):
-        thumb_update = False
-        significant_change = False
-        autotracker_update = False
-        # if the object is not in the current frame, add a 0.0 to the score history
-        if obj_data["frame_time"] != current_frame_time:
-            self.score_history.append(0.0)
-        else:
-            self.score_history.append(obj_data["score"])
-
-        # only keep the last 10 scores
-        if len(self.score_history) > 10:
-            self.score_history = self.score_history[-10:]
-
-        # calculate if this is a false positive
-        self.computed_score = self.compute_score()
-        if self.computed_score > self.top_score:
-            self.top_score = self.computed_score
-        self.false_positive = self._is_false_positive()
-        self.active = self.is_active()
-
-        if not self.false_positive and has_valid_frame:
-            # determine if this frame is a better thumbnail
-            if self.thumbnail_data is None or is_better_thumbnail(
-                self.obj_data["label"],
-                self.thumbnail_data,
-                obj_data,
-                self.camera_config.frame_shape,
-            ):
-                self.thumbnail_data = {
-                    "frame_time": current_frame_time,
-                    "box": obj_data["box"],
-                    "area": obj_data["area"],
-                    "region": obj_data["region"],
-                    "score": obj_data["score"],
-                    "attributes": obj_data["attributes"],
-                }
-                thumb_update = True
-
-        # check zones
-        current_zones = []
-        bottom_center = (obj_data["centroid"][0], obj_data["box"][3])
-        # check each zone
-        for name, zone in self.camera_config.zones.items():
-            # if the zone is not for this object type, skip
-            if len(zone.objects) > 0 and obj_data["label"] not in zone.objects:
-                continue
-            contour = zone.contour
-            zone_score = self.zone_presence.get(name, 0) + 1
-            # check if the object is in the zone
-            if cv2.pointPolygonTest(contour, bottom_center, False) >= 0:
-                # if the object passed the filters once, dont apply again
-                if name in self.current_zones or not zone_filtered(self, zone.filters):
-                    # an object is only considered present in a zone if it has a zone inertia of 3+
-                    if zone_score >= zone.inertia:
-                        loitering_score = self.zone_loitering.get(name, 0) + 1
-
-                        # loitering time is configured as seconds, convert to count of frames
-                        if loitering_score >= (
-                            self.camera_config.zones[name].loitering_time
-                            * self.camera_config.detect.fps
-                        ):
-                            current_zones.append(name)
-
-                            if name not in self.entered_zones:
-                                self.entered_zones.append(name)
-                        else:
-                            self.zone_loitering[name] = loitering_score
-                    else:
-                        self.zone_presence[name] = zone_score
-            else:
-                # once an object has a zone inertia of 3+ it is not checked anymore
-                if 0 < zone_score < zone.inertia:
-                    self.zone_presence[name] = zone_score - 1
-
-        # maintain attributes
-        for attr in obj_data["attributes"]:
-            if self.attributes[attr["label"]] < attr["score"]:
-                self.attributes[attr["label"]] = attr["score"]
-
-        # populate the sub_label for object with highest scoring logo
-        if self.obj_data["label"] in ["car", "package", "person"]:
-            recognized_logos = {
-                k: self.attributes[k]
-                for k in ["ups", "fedex", "amazon"]
-                if k in self.attributes
-            }
-            if len(recognized_logos) > 0:
-                max_logo = max(recognized_logos, key=recognized_logos.get)
-
-                # don't overwrite sub label if it is already set
-                if (
-                    self.obj_data.get("sub_label") is None
-                    or self.obj_data["sub_label"][0] == max_logo
-                ):
-                    self.obj_data["sub_label"] = (max_logo, recognized_logos[max_logo])
-
-        # check for significant change
-        if not self.false_positive:
-            # if the zones changed, signal an update
-            if set(self.current_zones) != set(current_zones):
-                significant_change = True
-
-            # if the position changed, signal an update
-            if self.obj_data["position_changes"] != obj_data["position_changes"]:
-                significant_change = True
-
-            if self.obj_data["attributes"] != obj_data["attributes"]:
-                significant_change = True
-
-            # if the state changed between stationary and active
-            if self.previous["active"] != self.active:
-                significant_change = True
-
-            # update at least once per minute
-            if self.obj_data["frame_time"] - self.previous["frame_time"] > 60:
-                significant_change = True
-
-            # update autotrack at most 3 objects per second
-            if self.obj_data["frame_time"] - self.previous["frame_time"] >= (1 / 3):
-                autotracker_update = True
-
-        self.obj_data.update(obj_data)
-        self.current_zones = current_zones
-        return (thumb_update, significant_change, autotracker_update)
-
-    def to_dict(self, include_thumbnail: bool = False):
-        event = {
-            "id": self.obj_data["id"],
-            "camera": self.camera,
-            "frame_time": self.obj_data["frame_time"],
-            "snapshot": self.thumbnail_data,
-            "label": self.obj_data["label"],
-            "sub_label": self.obj_data.get("sub_label"),
-            "top_score": self.top_score,
-            "false_positive": self.false_positive,
-            "start_time": self.obj_data["start_time"],
-            "end_time": self.obj_data.get("end_time", None),
-            "score": self.obj_data["score"],
-            "box": self.obj_data["box"],
-            "area": self.obj_data["area"],
-            "ratio": self.obj_data["ratio"],
-            "region": self.obj_data["region"],
-            "active": self.active,
-            "stationary": not self.active,
-            "motionless_count": self.obj_data["motionless_count"],
-            "position_changes": self.obj_data["position_changes"],
-            "current_zones": self.current_zones.copy(),
-            "entered_zones": self.entered_zones.copy(),
-            "has_clip": self.has_clip,
-            "has_snapshot": self.has_snapshot,
-            "attributes": self.attributes,
-            "current_attributes": self.obj_data["attributes"],
-        }
-
-        if include_thumbnail:
-            event["thumbnail"] = base64.b64encode(self.get_thumbnail()).decode("utf-8")
-
-        return event
-
-    def is_active(self):
-        return not self.is_stationary()
-
-    def is_stationary(self):
-        return (
-            self.obj_data["motionless_count"]
-            > self.camera_config.detect.stationary.threshold
-        )
-
-    def get_thumbnail(self):
-        if (
-            self.thumbnail_data is None
-            or self.thumbnail_data["frame_time"] not in self.frame_cache
-        ):
-            ret, jpg = cv2.imencode(".jpg", np.zeros((175, 175, 3), np.uint8))
-
-        jpg_bytes = self.get_jpg_bytes(
-            timestamp=False, bounding_box=False, crop=True, height=175
-        )
-
-        if jpg_bytes:
-            return jpg_bytes
-        else:
-            ret, jpg = cv2.imencode(".jpg", np.zeros((175, 175, 3), np.uint8))
-            return jpg.tobytes()
-
-    def get_clean_png(self):
-        if self.thumbnail_data is None:
-            return None
-
-        try:
-            best_frame = cv2.cvtColor(
-                self.frame_cache[self.thumbnail_data["frame_time"]],
-                cv2.COLOR_YUV2BGR_I420,
-            )
-        except KeyError:
-            logger.warning(
-                f"Unable to create clean png because frame {self.thumbnail_data['frame_time']} is not in the cache"
-            )
-            return None
-
-        ret, png = cv2.imencode(".png", best_frame)
-        if ret:
-            return png.tobytes()
-        else:
-            return None
-
-    def get_jpg_bytes(
-        self, timestamp=False, bounding_box=False, crop=False, height=None, quality=70
-    ):
-        if self.thumbnail_data is None:
-            return None
-
-        try:
-            best_frame = cv2.cvtColor(
-                self.frame_cache[self.thumbnail_data["frame_time"]],
-                cv2.COLOR_YUV2BGR_I420,
-            )
-        except KeyError:
-            logger.warning(
-                f"Unable to create jpg because frame {self.thumbnail_data['frame_time']} is not in the cache"
-            )
-            return None
-
-        if bounding_box:
-            thickness = 2
-            color = self.colormap[self.obj_data["label"]]
-
-            # draw the bounding boxes on the frame
-            box = self.thumbnail_data["box"]
-            draw_box_with_label(
-                best_frame,
-                box[0],
-                box[1],
-                box[2],
-                box[3],
-                self.obj_data["label"],
-                f"{int(self.thumbnail_data['score']*100)}% {int(self.thumbnail_data['area'])}",
-                thickness=thickness,
-                color=color,
-            )
-
-            # draw any attributes
-            for attribute in self.thumbnail_data["attributes"]:
-                box = attribute["box"]
-                draw_box_with_label(
-                    best_frame,
-                    box[0],
-                    box[1],
-                    box[2],
-                    box[3],
-                    attribute["label"],
-                    f"{attribute['score']:.0%}",
-                    thickness=thickness,
-                    color=color,
-                )
-
-        if crop:
-            box = self.thumbnail_data["box"]
-            box_size = 300
-            region = calculate_region(
-                best_frame.shape,
-                box[0],
-                box[1],
-                box[2],
-                box[3],
-                box_size,
-                multiplier=1.1,
-            )
-            best_frame = best_frame[region[1] : region[3], region[0] : region[2]]
-
-        if height:
-            width = int(height * best_frame.shape[1] / best_frame.shape[0])
-            best_frame = cv2.resize(
-                best_frame, dsize=(width, height), interpolation=cv2.INTER_AREA
-            )
-        if timestamp:
-            color = self.camera_config.timestamp_style.color
-            draw_timestamp(
-                best_frame,
-                self.thumbnail_data["frame_time"],
-                self.camera_config.timestamp_style.format,
-                font_effect=self.camera_config.timestamp_style.effect,
-                font_thickness=self.camera_config.timestamp_style.thickness,
-                font_color=(color.blue, color.green, color.red),
-                position=self.camera_config.timestamp_style.position,
-            )
-
-        ret, jpg = cv2.imencode(
-            ".jpg", best_frame, [int(cv2.IMWRITE_JPEG_QUALITY), quality]
-        )
-        if ret:
-            return jpg.tobytes()
-        else:
-            return None
-
-
-def zone_filtered(obj: TrackedObject, object_config):
-    object_name = obj.obj_data["label"]
-
-    if object_name in object_config:
-        obj_settings = object_config[object_name]
-
-        # if the min area is larger than the
-        # detected object, don't add it to detected objects
-        if obj_settings.min_area > obj.obj_data["area"]:
-            return True
-
-        # if the detected object is larger than the
-        # max area, don't add it to detected objects
-        if obj_settings.max_area < obj.obj_data["area"]:
-            return True
-
-        # if the score is lower than the threshold, skip
-        if obj_settings.threshold > obj.computed_score:
-            return True
-
-        # if the object is not proportionally wide enough
-        if obj_settings.min_ratio > obj.obj_data["ratio"]:
-            return True
-
-        # if the object is proportionally too wide
-        if obj_settings.max_ratio < obj.obj_data["ratio"]:
-            return True
-
-    return False
-
-
 # Maintains the state of a camera
 class CameraState:
     def __init__(
@@ -676,17 +233,18 @@ class CameraState:
     def on(self, event_type: str, callback: Callable[[dict], None]):
         self.callbacks[event_type].append(callback)
 
-    def update(self, frame_time, current_detections, motion_boxes, regions):
-        # get the new frame
-        frame_id = f"{self.name}{frame_time}"
-
+    def update(
+        self,
+        frame_name: str,
+        frame_time: float,
+        current_detections: dict[str, dict[str, any]],
+        motion_boxes: list[tuple[int, int, int, int]],
+        regions: list[tuple[int, int, int, int]],
+    ):
         current_frame = self.frame_manager.get(
-            frame_id, self.camera_config.frame_shape_yuv
+            frame_name, self.camera_config.frame_shape_yuv
         )
 
-        if current_frame is None:
-            logger.debug(f"Failed to get frame {frame_id} from SHM")
-
         tracked_objects = self.tracked_objects.copy()
         current_ids = set(current_detections.keys())
         previous_ids = set(tracked_objects.keys())
@@ -696,8 +254,7 @@ class CameraState:
 
         for id in new_ids:
             new_obj = tracked_objects[id] = TrackedObject(
-                self.name,
-                self.config.model.colormap,
+                self.config.model,
                 self.camera_config,
                 self.frame_cache,
                 current_detections[id],
@@ -705,7 +262,7 @@ class CameraState:
 
             # call event handlers
             for c in self.callbacks["start"]:
-                c(self.name, new_obj, frame_time)
+                c(self.name, new_obj, frame_name)
 
         for id in updated_ids:
             updated_obj = tracked_objects[id]
@@ -715,7 +272,7 @@ class CameraState:
 
             if autotracker_update or significant_update:
                 for c in self.callbacks["autotrack"]:
-                    c(self.name, updated_obj, frame_time)
+                    c(self.name, updated_obj, frame_name)
 
             if thumb_update and current_frame is not None:
                 # ensure this frame is stored in the cache
@@ -736,7 +293,7 @@ class CameraState:
             ) or significant_update:
                 # call event handlers
                 for c in self.callbacks["update"]:
-                    c(self.name, updated_obj, frame_time)
+                    c(self.name, updated_obj, frame_name)
                 updated_obj.last_published = frame_time
 
         for id in removed_ids:
@@ -745,7 +302,7 @@ class CameraState:
             if "end_time" not in removed_obj.obj_data:
                 removed_obj.obj_data["end_time"] = frame_time
                 for c in self.callbacks["end"]:
-                    c(self.name, removed_obj, frame_time)
+                    c(self.name, removed_obj, frame_name)
 
         # TODO: can i switch to looking this up and only changing when an event ends?
         # maintain best objects
@@ -788,6 +345,7 @@ class CameraState:
             # if the object's thumbnail is not from the current frame, skip
             if (
                 current_frame is None
+                or obj.thumbnail_data is None
                 or obj.false_positive
                 or obj.thumbnail_data["frame_time"] != frame_time
             ):
@@ -810,11 +368,11 @@ class CameraState:
                 ):
                     self.best_objects[object_type] = obj
                     for c in self.callbacks["snapshot"]:
-                        c(self.name, self.best_objects[object_type], frame_time)
+                        c(self.name, self.best_objects[object_type], frame_name)
             else:
                 self.best_objects[object_type] = obj
                 for c in self.callbacks["snapshot"]:
-                    c(self.name, self.best_objects[object_type], frame_time)
+                    c(self.name, self.best_objects[object_type], frame_name)
 
         for c in self.callbacks["camera_activity"]:
             c(self.name, camera_activity)
@@ -889,7 +447,7 @@ class CameraState:
                     c(self.name, obj_name, 0)
                 self.active_object_counts[obj_name] = 0
             for c in self.callbacks["snapshot"]:
-                c(self.name, self.best_objects[obj_name], frame_time)
+                c(self.name, self.best_objects[obj_name], frame_name)
 
         # cleanup thumbnail frame cache
         current_thumb_frames = {
@@ -920,7 +478,7 @@ class CameraState:
                 if self.previous_frame_id is not None:
                     self.frame_manager.close(self.previous_frame_id)
 
-            self.previous_frame_id = frame_id
+            self.previous_frame_id = frame_name
 
 
 class TrackedObjectProcessor(threading.Thread):
@@ -960,17 +518,18 @@ class TrackedObjectProcessor(threading.Thread):
         self.zone_data = defaultdict(lambda: defaultdict(dict))
         self.active_zone_data = defaultdict(lambda: defaultdict(dict))
 
-        def start(camera, obj: TrackedObject, current_frame_time):
+        def start(camera: str, obj: TrackedObject, frame_name: str):
             self.event_sender.publish(
                 (
                     EventTypeEnum.tracked_object,
                     EventStateEnum.start,
                     camera,
+                    frame_name,
                     obj.to_dict(),
                 )
             )
 
-        def update(camera, obj: TrackedObject, current_frame_time):
+        def update(camera: str, obj: TrackedObject, frame_name: str):
             obj.has_snapshot = self.should_save_snapshot(camera, obj)
             obj.has_clip = self.should_retain_recording(camera, obj)
             after = obj.to_dict()
@@ -986,14 +545,15 @@ class TrackedObjectProcessor(threading.Thread):
                     EventTypeEnum.tracked_object,
                     EventStateEnum.update,
                     camera,
+                    frame_name,
                     obj.to_dict(include_thumbnail=True),
                 )
             )
 
-        def autotrack(camera, obj: TrackedObject, current_frame_time):
+        def autotrack(camera: str, obj: TrackedObject, frame_name: str):
             self.ptz_autotracker_thread.ptz_autotracker.autotrack_object(camera, obj)
 
-        def end(camera, obj: TrackedObject, current_frame_time):
+        def end(camera: str, obj: TrackedObject, frame_name: str):
             # populate has_snapshot
             obj.has_snapshot = self.should_save_snapshot(camera, obj)
             obj.has_clip = self.should_retain_recording(camera, obj)
@@ -1048,11 +608,12 @@ class TrackedObjectProcessor(threading.Thread):
                     EventTypeEnum.tracked_object,
                     EventStateEnum.end,
                     camera,
+                    frame_name,
                     obj.to_dict(include_thumbnail=True),
                 )
             )
 
-        def snapshot(camera, obj: TrackedObject, current_frame_time):
+        def snapshot(camera, obj: TrackedObject, frame_name: str):
             mqtt_config: MqttConfig = self.config.cameras[camera].mqtt
             if mqtt_config.enabled and self.should_mqtt_snapshot(camera, obj):
                 jpg_bytes = obj.get_jpg_bytes(
@@ -1157,7 +718,8 @@ class TrackedObjectProcessor(threading.Thread):
                 )
                 and (
                     not review_config.detections.required_zones
-                    or set(obj.entered_zones) & set(review_config.alerts.required_zones)
+                    or set(obj.entered_zones)
+                    & set(review_config.detections.required_zones)
                 )
             )
         ):
@@ -1240,6 +802,7 @@ class TrackedObjectProcessor(threading.Thread):
             try:
                 (
                     camera,
+                    frame_name,
                     frame_time,
                     current_tracked_objects,
                     motion_boxes,
@@ -1251,7 +814,7 @@ class TrackedObjectProcessor(threading.Thread):
             camera_state = self.camera_states[camera]
 
             camera_state.update(
-                frame_time, current_tracked_objects, motion_boxes, regions
+                frame_name, frame_time, current_tracked_objects, motion_boxes, regions
             )
 
             self.update_mqtt_motion(camera, frame_time, motion_boxes)
@@ -1264,6 +827,7 @@ class TrackedObjectProcessor(threading.Thread):
             self.detection_publisher.publish(
                 (
                     camera,
+                    frame_name,
                     frame_time,
                     tracked_objects,
                     motion_boxes,

frigate/output/birdseye.py

@@ -268,12 +268,10 @@ class BirdsEyeFrameManager:
     def __init__(
         self,
         config: FrigateConfig,
-        frame_manager: SharedMemoryFrameManager,
         stop_event: mp.Event,
     ):
         self.config = config
         self.mode = config.birdseye.mode
-        self.frame_manager = frame_manager
         width, height = get_canvas_shape(config.birdseye.width, config.birdseye.height)
         self.frame_shape = (height, width)
         self.yuv_shape = (height * 3 // 2, width)
@@ -351,18 +349,13 @@ class BirdsEyeFrameManager:
         logger.debug("Clearing the birdseye frame")
         self.frame[:] = self.blank_frame
 
-    def copy_to_position(self, position, camera=None, frame_time=None):
+    def copy_to_position(self, position, camera=None, frame: np.ndarray = None):
         if camera is None:
             frame = None
             channel_dims = None
         else:
-            frame_id = f"{camera}{frame_time}"
-            frame = self.frame_manager.get(
-                frame_id, self.config.cameras[camera].frame_shape_yuv
-            )
-
             if frame is None:
-                logger.debug(f"Unable to copy frame {camera}{frame_time} to birdseye.")
+                logger.debug(f"Unable to copy frame {camera} to birdseye.")
                 return
 
             channel_dims = self.cameras[camera]["channel_dims"]
@@ -375,8 +368,6 @@ class BirdsEyeFrameManager:
             channel_dims,
         )
 
-        self.frame_manager.close(frame_id)
-
     def camera_active(self, mode, object_box_count, motion_box_count):
         if mode == BirdseyeModeEnum.continuous:
             return True
@@ -387,7 +378,7 @@ class BirdsEyeFrameManager:
         if mode == BirdseyeModeEnum.objects and object_box_count > 0:
             return True
 
-    def update_frame(self):
+    def update_frame(self, frame: np.ndarray):
         """Update to a new frame for birdseye."""
 
         # determine how many cameras are tracking objects within the last inactivity_threshold seconds
@@ -397,7 +388,7 @@ class BirdsEyeFrameManager:
                 for cam, cam_data in self.cameras.items()
                 if self.config.cameras[cam].birdseye.enabled
                 and cam_data["last_active_frame"] > 0
-                and cam_data["current_frame"] - cam_data["last_active_frame"]
+                and cam_data["current_frame_time"] - cam_data["last_active_frame"]
                 < self.inactivity_threshold
             ]
         )
@@ -414,7 +405,7 @@ class BirdsEyeFrameManager:
                 limited_active_cameras = sorted(
                     active_cameras,
                     key=lambda active_camera: (
-                        self.cameras[active_camera]["current_frame"]
+                        self.cameras[active_camera]["current_frame_time"]
                         - self.cameras[active_camera]["last_active_frame"]
                     ),
                 )
@@ -524,7 +515,9 @@ class BirdsEyeFrameManager:
         for row in self.camera_layout:
             for position in row:
                 self.copy_to_position(
-                    position[1], position[0], self.cameras[position[0]]["current_frame"]
+                    position[1],
+                    position[0],
+                    self.cameras[position[0]]["current_frame"],
                 )
 
         return True
@@ -672,7 +665,14 @@ class BirdsEyeFrameManager:
         else:
             return standard_candidate_layout
 
-    def update(self, camera, object_count, motion_count, frame_time, frame) -> bool:
+    def update(
+        self,
+        camera: str,
+        object_count: int,
+        motion_count: int,
+        frame_time: float,
+        frame: np.ndarray,
+    ) -> bool:
         # don't process if birdseye is disabled for this camera
         camera_config = self.config.cameras[camera].birdseye
 
@@ -689,7 +689,8 @@ class BirdsEyeFrameManager:
             return False
 
         # update the last active frame for the camera
-        self.cameras[camera]["current_frame"] = frame_time
+        self.cameras[camera]["current_frame"] = frame.copy()
+        self.cameras[camera]["current_frame_time"] = frame_time
         if self.camera_active(camera_config.mode, object_count, motion_count):
             self.cameras[camera]["last_active_frame"] = frame_time
 
@@ -700,7 +701,7 @@ class BirdsEyeFrameManager:
             return False
 
         try:
-            updated_frame = self.update_frame()
+            updated_frame = self.update_frame(frame)
         except Exception:
             updated_frame = False
             self.active_cameras = []
@@ -737,12 +738,12 @@ class Birdseye:
         self.broadcaster = BroadcastThread(
             "birdseye", self.converter, websocket_server, stop_event
         )
-        frame_manager = SharedMemoryFrameManager()
-        self.birdseye_manager = BirdsEyeFrameManager(config, frame_manager, stop_event)
+        self.birdseye_manager = BirdsEyeFrameManager(config, stop_event)
         self.config_subscriber = ConfigSubscriber("config/birdseye/")
+        self.frame_manager = SharedMemoryFrameManager()
 
         if config.birdseye.restream:
-            self.birdseye_buffer = frame_manager.create(
+            self.birdseye_buffer = self.frame_manager.create(
                 "birdseye",
                 self.birdseye_manager.yuv_shape[0] * self.birdseye_manager.yuv_shape[1],
             )
@@ -756,7 +757,7 @@ class Birdseye:
         current_tracked_objects: list[dict[str, any]],
         motion_boxes: list[list[int]],
         frame_time: float,
-        frame,
+        frame: np.ndarray,
     ) -> None:
         # check if there is an updated config
         while True:

frigate/output/output.py

@@ -63,6 +63,7 @@ def output_frames(
     birdseye: Optional[Birdseye] = None
     preview_recorders: dict[str, PreviewRecorder] = {}
     preview_write_times: dict[str, float] = {}
+    failed_frame_requests: dict[str, int] = {}
 
     move_preview_frames("cache")
 
@@ -87,19 +88,27 @@ def output_frames(
 
         (
             camera,
+            frame_name,
             frame_time,
             current_tracked_objects,
             motion_boxes,
-            regions,
+            _,
         ) = data
 
-        frame_id = f"{camera}{frame_time}"
-
-        frame = frame_manager.get(frame_id, config.cameras[camera].frame_shape_yuv)
+        frame = frame_manager.get(frame_name, config.cameras[camera].frame_shape_yuv)
 
         if frame is None:
-            logger.debug(f"Failed to get frame {frame_id} from SHM")
+            logger.debug(f"Failed to get frame {frame_name} from SHM")
+            failed_frame_requests[camera] = failed_frame_requests.get(camera, 0) + 1
+
+            if failed_frame_requests[camera] > config.cameras[camera].detect.fps:
+                logger.warning(
+                    f"Failed to retrieve many frames for {camera} from SHM, consider increasing SHM size if this continues."
+                )
+
             continue
+        else:
+            failed_frame_requests[camera] = 0
 
         # send camera frame to ffmpeg process if websockets are connected
         if any(
@@ -134,12 +143,15 @@ def output_frames(
         # check for any cameras that are currently offline
         # and need to generate a preview
         if generated_preview:
+            logger.debug(
+                "Checking for offline cameras because another camera generated a preview."
+            )
             for camera, time in preview_write_times.copy().items():
                 if time != 0 and frame_time - time > 10:
                     preview_recorders[camera].flag_offline(frame_time)
                     preview_write_times[camera] = frame_time
 
-        frame_manager.close(frame_id)
+        frame_manager.close(frame_name)
 
     move_preview_frames("clips")
 
@@ -151,15 +163,15 @@ def output_frames(
 
         (
             camera,
+            frame_name,
             frame_time,
             current_tracked_objects,
             motion_boxes,
             regions,
         ) = data
 
-        frame_id = f"{camera}{frame_time}"
-        frame = frame_manager.get(frame_id, config.cameras[camera].frame_shape_yuv)
-        frame_manager.close(frame_id)
+        frame = frame_manager.get(frame_name, config.cameras[camera].frame_shape_yuv)
+        frame_manager.close(frame_name)
 
     detection_subscriber.stop()
 

frigate/output/preview.py

@@ -78,7 +78,7 @@ class FFMpegConverter(threading.Thread):
         # write a PREVIEW at fps and 1 key frame per clip
         self.ffmpeg_cmd = parse_preset_hardware_acceleration_encode(
             config.ffmpeg.ffmpeg_path,
-            config.ffmpeg.hwaccel_args,
+            "default",
             input="-f concat -y -protocol_whitelist pipe,file -safe 0 -threads 1 -i /dev/stdin",
             output=f"-threads 1 -g {PREVIEW_KEYFRAME_INTERVAL} -bf 0 -b:v {PREVIEW_QUALITY_BIT_RATES[self.config.record.preview.quality]} {FPS_VFR_PARAM} -movflags +faststart -pix_fmt yuv420p {self.path}",
             type=EncodeTypeEnum.preview,
@@ -154,6 +154,7 @@ class PreviewRecorder:
         self.start_time = 0
         self.last_output_time = 0
         self.output_frames = []
+
         if config.detect.width > config.detect.height:
             self.out_height = PREVIEW_HEIGHT
             self.out_width = (
@@ -274,7 +275,7 @@ class PreviewRecorder:
 
         return False
 
-    def write_frame_to_cache(self, frame_time: float, frame) -> None:
+    def write_frame_to_cache(self, frame_time: float, frame: np.ndarray) -> None:
         # resize yuv frame
         small_frame = np.zeros((self.out_height * 3 // 2, self.out_width), np.uint8)
         copy_yuv_to_position(
@@ -303,7 +304,7 @@ class PreviewRecorder:
         current_tracked_objects: list[dict[str, any]],
         motion_boxes: list[list[int]],
         frame_time: float,
-        frame,
+        frame: np.ndarray,
     ) -> bool:
         # check for updated record config
         _, updated_record_config = self.config_subscriber.check_for_update()
@@ -332,6 +333,10 @@ class PreviewRecorder:
                     self.output_frames,
                     self.requestor,
                 ).start()
+            else:
+                logger.debug(
+                    f"Not saving preview for {self.config.name} because there are no saved frames."
+                )
 
             # reset frame cache
             self.segment_end = (

frigate/ptz/autotrack.py

@@ -32,6 +32,7 @@ from frigate.const import (
     CONFIG_DIR,
 )
 from frigate.ptz.onvif import OnvifController
+from frigate.track.tracked_object import TrackedObject
 from frigate.util.builtin import update_yaml_file
 from frigate.util.image import SharedMemoryFrameManager, intersection_over_union
 
@@ -58,7 +59,13 @@ class PtzMotionEstimator:
         self.ptz_metrics.reset.set()
         logger.debug(f"{config.name}: Motion estimator init")
 
-    def motion_estimator(self, detections, frame_time, camera):
+    def motion_estimator(
+        self,
+        detections: list[dict[str, any]],
+        frame_name: str,
+        frame_time: float,
+        camera: str,
+    ):
         # If we've just started up or returned to our preset, reset motion estimator for new tracking session
         if self.ptz_metrics.reset.is_set():
             self.ptz_metrics.reset.clear()
@@ -91,9 +98,8 @@ class PtzMotionEstimator:
                 f"{camera}: Motion estimator running - frame time: {frame_time}"
             )
 
-            frame_id = f"{camera}{frame_time}"
             yuv_frame = self.frame_manager.get(
-                frame_id, self.camera_config.frame_shape_yuv
+                frame_name, self.camera_config.frame_shape_yuv
             )
 
             if yuv_frame is None:
@@ -135,7 +141,7 @@ class PtzMotionEstimator:
             except Exception:
                 pass
 
-            self.frame_manager.close(frame_id)
+            self.frame_manager.close(frame_name)
 
         return self.coord_transformations
 
@@ -214,7 +220,7 @@ class PtzAutoTracker:
             ):
                 self._autotracker_setup(camera_config, camera)
 
-    def _autotracker_setup(self, camera_config, camera):
+    def _autotracker_setup(self, camera_config: CameraConfig, camera: str):
         logger.debug(f"{camera}: Autotracker init")
 
         self.object_types[camera] = camera_config.onvif.autotracking.track
@@ -852,7 +858,7 @@ class PtzAutoTracker:
             logger.debug(f"{camera}: Valid velocity ")
             return True, velocities.flatten()
 
-    def _get_distance_threshold(self, camera, obj):
+    def _get_distance_threshold(self, camera: str, obj: TrackedObject):
         # Returns true if Euclidean distance from object to center of frame is
         # less than 10% of the of the larger dimension (width or height) of the frame,
         # multiplied by a scaling factor for object size.
@@ -888,7 +894,9 @@ class PtzAutoTracker:
 
         return distance_threshold
 
-    def _should_zoom_in(self, camera, obj, box, predicted_time, debug_zooming=False):
+    def _should_zoom_in(
+        self, camera: str, obj: TrackedObject, box, predicted_time, debug_zooming=False
+    ):
         # returns True if we should zoom in, False if we should zoom out, None to do nothing
         camera_config = self.config.cameras[camera]
         camera_width = camera_config.frame_shape[1]
@@ -1019,7 +1027,7 @@ class PtzAutoTracker:
         # Don't zoom at all
         return None
 
-    def _autotrack_move_ptz(self, camera, obj):
+    def _autotrack_move_ptz(self, camera: str, obj: TrackedObject):
         camera_config = self.config.cameras[camera]
         camera_width = camera_config.frame_shape[1]
         camera_height = camera_config.frame_shape[0]
@@ -1090,7 +1098,12 @@ class PtzAutoTracker:
                 self._enqueue_move(camera, obj.obj_data["frame_time"], 0, 0, zoom)
 
     def _get_zoom_amount(
-        self, camera, obj, predicted_box, predicted_movement_time, debug_zoom=True
+        self,
+        camera: str,
+        obj: TrackedObject,
+        predicted_box,
+        predicted_movement_time,
+        debug_zoom=True,
     ):
         camera_config = self.config.cameras[camera]
 
@@ -1186,13 +1199,13 @@ class PtzAutoTracker:
 
         return zoom
 
-    def is_autotracking(self, camera):
+    def is_autotracking(self, camera: str):
         return self.tracked_object[camera] is not None
 
-    def autotracked_object_region(self, camera):
+    def autotracked_object_region(self, camera: str):
         return self.tracked_object[camera]["region"]
 
-    def autotrack_object(self, camera, obj):
+    def autotrack_object(self, camera: str, obj: TrackedObject):
         camera_config = self.config.cameras[camera]
 
         if camera_config.onvif.autotracking.enabled:
@@ -1208,7 +1221,7 @@ class PtzAutoTracker:
             if (
                 # new object
                 self.tracked_object[camera] is None
-                and obj.camera == camera
+                and obj.camera_config.name == camera
                 and obj.obj_data["label"] in self.object_types[camera]
                 and set(obj.entered_zones) & set(self.required_zones[camera])
                 and not obj.previous["false_positive"]
@@ -1267,7 +1280,7 @@ class PtzAutoTracker:
                 # If it's within bounds, start tracking that object.
                 # Should we check region (maybe too broad) or expand the previous object's box a bit and check that?
                 self.tracked_object[camera] is None
-                and obj.camera == camera
+                and obj.camera_config.name == camera
                 and obj.obj_data["label"] in self.object_types[camera]
                 and not obj.previous["false_positive"]
                 and not obj.false_positive

frigate/record/export.py

@@ -27,6 +27,7 @@ from frigate.ffmpeg_presets import (
     parse_preset_hardware_acceleration_encode,
 )
 from frigate.models import Export, Previews, Recordings
+from frigate.util.builtin import is_current_hour
 
 logger = logging.getLogger(__name__)
 
@@ -43,6 +44,11 @@ class PlaybackFactorEnum(str, Enum):
     timelapse_25x = "timelapse_25x"
 
 
+class PlaybackSourceEnum(str, Enum):
+    recordings = "recordings"
+    preview = "preview"
+
+
 class RecordingExporter(threading.Thread):
     """Exports a specific set of recordings for a camera to storage as a single file."""
 
@@ -56,6 +62,7 @@ class RecordingExporter(threading.Thread):
         start_time: int,
         end_time: int,
         playback_factor: PlaybackFactorEnum,
+        playback_source: PlaybackSourceEnum,
     ) -> None:
         super().__init__()
         self.config = config
@@ -66,6 +73,7 @@ class RecordingExporter(threading.Thread):
         self.start_time = start_time
         self.end_time = end_time
         self.playback_factor = playback_factor
+        self.playback_source = playback_source
 
         # ensure export thumb dir
         Path(os.path.join(CLIPS_DIR, "export")).mkdir(exist_ok=True)
@@ -170,30 +178,7 @@ class RecordingExporter(threading.Thread):
 
         return thumb_path
 
-    def run(self) -> None:
-        logger.debug(
-            f"Beginning export for {self.camera} from {self.start_time} to {self.end_time}"
-        )
-        export_name = (
-            self.user_provided_name
-            or f"{self.camera.replace('_', ' ')} {self.get_datetime_from_timestamp(self.start_time)} {self.get_datetime_from_timestamp(self.end_time)}"
-        )
-        video_path = f"{EXPORT_DIR}/{self.export_id}.mp4"
-
-        thumb_path = self.save_thumbnail(self.export_id)
-
-        Export.insert(
-            {
-                Export.id: self.export_id,
-                Export.camera: self.camera,
-                Export.name: export_name,
-                Export.date: self.start_time,
-                Export.video_path: video_path,
-                Export.thumb_path: thumb_path,
-                Export.in_progress: True,
-            }
-        ).execute()
-
+    def get_record_export_command(self, video_path: str) -> list[str]:
         if (self.end_time - self.start_time) <= MAX_PLAYLIST_SECONDS:
             playlist_lines = f"http://127.0.0.1:5000/vod/{self.camera}/start/{self.start_time}/end/{self.end_time}/index.m3u8"
             ffmpeg_input = (
@@ -204,7 +189,10 @@ class RecordingExporter(threading.Thread):
 
             # get full set of recordings
             export_recordings = (
-                Recordings.select()
+                Recordings.select(
+                    Recordings.start_time,
+                    Recordings.end_time,
+                )
                 .where(
                     Recordings.start_time.between(self.start_time, self.end_time)
                     | Recordings.end_time.between(self.start_time, self.end_time)
@@ -233,6 +221,91 @@ class RecordingExporter(threading.Thread):
             ffmpeg_cmd = (
                 f"{self.config.ffmpeg.ffmpeg_path} -hide_banner {ffmpeg_input} -c copy -movflags +faststart {video_path}"
             ).split(" ")
+        elif self.playback_factor == PlaybackFactorEnum.timelapse_25x:
+            ffmpeg_cmd = (
+                parse_preset_hardware_acceleration_encode(
+                    self.config.ffmpeg.ffmpeg_path,
+                    self.config.ffmpeg.hwaccel_args,
+                    f"-an {ffmpeg_input}",
+                    f"{self.config.cameras[self.camera].record.export.timelapse_args} -movflags +faststart {video_path}",
+                    EncodeTypeEnum.timelapse,
+                )
+            ).split(" ")
+
+        return ffmpeg_cmd, playlist_lines
+
+    def get_preview_export_command(self, video_path: str) -> list[str]:
+        playlist_lines = []
+        codec = "-c copy"
+
+        if is_current_hour(self.start_time):
+            # get list of current preview frames
+            preview_dir = os.path.join(CACHE_DIR, "preview_frames")
+            file_start = f"preview_{self.camera}"
+            start_file = f"{file_start}-{self.start_time}.{PREVIEW_FRAME_TYPE}"
+            end_file = f"{file_start}-{self.end_time}.{PREVIEW_FRAME_TYPE}"
+
+            for file in sorted(os.listdir(preview_dir)):
+                if not file.startswith(file_start):
+                    continue
+
+                if file < start_file:
+                    continue
+
+                if file > end_file:
+                    break
+
+                playlist_lines.append(f"file '{os.path.join(preview_dir, file)}'")
+                playlist_lines.append("duration 0.12")
+
+            if playlist_lines:
+                last_file = playlist_lines[-2]
+                playlist_lines.append(last_file)
+                codec = "-c:v libx264"
+
+        # get full set of previews
+        export_previews = (
+            Previews.select(
+                Previews.path,
+                Previews.start_time,
+                Previews.end_time,
+            )
+            .where(
+                Previews.start_time.between(self.start_time, self.end_time)
+                | Previews.end_time.between(self.start_time, self.end_time)
+                | (
+                    (self.start_time > Previews.start_time)
+                    & (self.end_time < Previews.end_time)
+                )
+            )
+            .where(Previews.camera == self.camera)
+            .order_by(Previews.start_time.asc())
+            .namedtuples()
+            .iterator()
+        )
+
+        preview: Previews
+        for preview in export_previews:
+            playlist_lines.append(f"file '{preview.path}'")
+
+            if preview.start_time < self.start_time:
+                playlist_lines.append(
+                    f"inpoint {int(self.start_time - preview.start_time)}"
+                )
+
+            if preview.end_time > self.end_time:
+                playlist_lines.append(
+                    f"outpoint {int(preview.end_time - self.end_time)}"
+                )
+
+        ffmpeg_input = (
+            "-y -protocol_whitelist pipe,file,tcp -f concat -safe 0 -i /dev/stdin"
+        )
+
+        if self.playback_factor == PlaybackFactorEnum.realtime:
+            ffmpeg_cmd = (
+                f"{self.config.ffmpeg.ffmpeg_path} -hide_banner {ffmpeg_input} {codec} -movflags +faststart {video_path}"
+            ).split(" ")
         elif self.playback_factor == PlaybackFactorEnum.timelapse_25x:
             ffmpeg_cmd = (
                 parse_preset_hardware_acceleration_encode(
@@ -244,6 +317,36 @@ class RecordingExporter(threading.Thread):
                 )
             ).split(" ")
 
+        return ffmpeg_cmd, playlist_lines
+
+    def run(self) -> None:
+        logger.debug(
+            f"Beginning export for {self.camera} from {self.start_time} to {self.end_time}"
+        )
+        export_name = (
+            self.user_provided_name
+            or f"{self.camera.replace('_', ' ')} {self.get_datetime_from_timestamp(self.start_time)} {self.get_datetime_from_timestamp(self.end_time)}"
+        )
+        video_path = f"{EXPORT_DIR}/{self.export_id}.mp4"
+        thumb_path = self.save_thumbnail(self.export_id)
+
+        Export.insert(
+            {
+                Export.id: self.export_id,
+                Export.camera: self.camera,
+                Export.name: export_name,
+                Export.date: self.start_time,
+                Export.video_path: video_path,
+                Export.thumb_path: thumb_path,
+                Export.in_progress: True,
+            }
+        ).execute()
+
+        if self.playback_source == PlaybackSourceEnum.recordings:
+            ffmpeg_cmd, playlist_lines = self.get_record_export_command(video_path)
+        else:
+            ffmpeg_cmd, playlist_lines = self.get_preview_export_command(video_path)
+
         p = sp.run(
             ffmpeg_cmd,
             input="\n".join(playlist_lines),
@@ -254,7 +357,7 @@ class RecordingExporter(threading.Thread):
 
         if p.returncode != 0:
             logger.error(
-                f"Failed to export recording for command {' '.join(ffmpeg_cmd)}"
+                f"Failed to export {self.playback_source.value} for command {' '.join(ffmpeg_cmd)}"
             )
             logger.error(p.stderr)
             Path(video_path).unlink(missing_ok=True)