Object detection is what allows Frigate to identify _what_ is in your camera's view (people, cars, animals, and more) rather than just reacting to pixel changes. When Frigate's motion detection finds activity in a frame, that region is sent to an **object detector**, which returns the objects it recognizes along with their location and a confidence score. These detections are what drive tracked objects, alerts, detections, and notifications.
Object detection is computationally intensive, so Frigate is designed to run it on a dedicated AI accelerator or GPU rather than the CPU. A **detector** is the specific hardware-and-model backend Frigate uses to run inference. Choosing a detector that matches your hardware is one of the most important steps in getting good performance, and the right choice depends on what device Frigate is running on.
- [Coral EdgeTPU](#edge-tpu-detector): The Google Coral EdgeTPU is available in USB, Mini PCIe, and m.2 formats allowing for a wide range of compatibility with devices.
- [Hailo](#hailo): The Hailo-8, Hailo-8L and Hailo-8R AI Acceleration modules are available in m.2 format with a HAT for RPi devices, offering a wide range of compatibility with devices.
- <CommunityBadge /> [MemryX](#memryx-mx3): The MX3 Acceleration module is available in m.2 format, offering broad compatibility across various platforms.
- <CommunityBadge /> [DEEPX](#deepx-npu): The DEEPX NPU is available in m.2 format and as a HAT+ for the Raspberry Pi 5, offering broad compatibility across various platforms.
- <CommunityBadge /> [XDNA2](#amd-xdna2): AMD Ryzen AI / XDNA2 NPUs can run object detection through the community-maintained `frigate-xdna` ZMQ sidecar.
- <CommunityBadge /> [ONNX](#apple-neural-engine-lighter): the ONNX detector runs on the Neural Engine of M1 and newer Macs when Frigate runs under the lighter container runtime.
- [OpenVino](#openvino-detector): OpenVino can run on Intel Arc GPUs, Intel integrated GPUs, and Intel CPUs to provide efficient object detection.
- [ONNX](#onnx): OpenVINO will automatically be detected and used as a detector in the default Frigate image when a supported ONNX model is configured.
- [ONNX](#onnx): Nvidia GPUs 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 Jetson devices, using one of many default models.
- [ONNX](#onnx): TensorRT will automatically be detected and used as a detector in the `-tensorrt-jp6` Frigate image when a supported ONNX model is configured.
- [Axelera Metis](#axelera-metis): The Metis AI Accelerator is available in M.2 and PCIe formats, running compiled YOLO models through the vendor runtime.
- [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.
A single model can not be spread across different detector types (ex: OpenVINO and Coral EdgeTPU can not run the same model at the same time). Configuring more than one model, each on its own detector type, is supported.
Object detection is configured with a `models` list. Each entry describes one model and the hardware it runs on:
```yaml
models:
- devices:
- openvino:GPU
path:/config/model_cache/yolov9-s.onnx
model_type:yolo-generic
width:320
height:320
```
Each entry in `devices` is a detector type, optionally followed by a colon and a device for that detector, such as `edgetpu:pci:0`, `openvino:NPU`, or `tensorrt:0`. The per-detector sections below document the device values each one accepts. Listing several devices runs the model on all of them, and listing the **same** device more than once runs additional inference processes against it, which can improve throughput on hardware that keeps up with more than one stream:
Cameras can be split across models by scene, which is useful when some cameras benefit from a differently trained model, such as thermal cameras. Each model declares the `scene` it is for, and each camera picks one with `detect -> scene`:
A scene is any name made up of letters, numbers, `_`, and `-`. The model with a scene of `default` is used by every camera that does not set one (or sets a scene that no model is configured for), and `default` is used when a model does not declare a scene. Changing a camera's scene requires a restart.
:::warning
Scenes are for running **different** models. Do not configure the same model under several scenes to dedicate a detector to specific cameras: every detector of a model already serves every camera using it, and splitting them only leaves some detectors idle while others fall behind. Frigate detects models that use the same model file, even under a different path or file name, combines them into one model with all of their hardware, and logs a warning.
Along with picking a detector for your hardware, you will choose a model's **input resolution** (such as `320x320` or `640x640`) and, for model families like YOLOv9, a **variant size** (`tiny`, `small`, etc.). Both affect the balance between accuracy and the inference time your hardware can sustain.
**Resolution (320x320 vs 640x640):** Frigate is optimized for `320x320` models, and `320x320` is the best choice for the vast majority of setups. Frigate is specifically designed to compensate for the smaller model by cropping a region of motion from the full frame and zooming into it before running detection, so a `320x320` model is actually _better_ at small and distant objects, not worse. A `640x640` model is slower and uses more resources, and its main benefit is fitting more objects into a single inference when many objects are spread across a large area. Recent versions of Frigate have improved support for `640x640` models, but `320x320` remains the recommended starting point for nearly all setups.
**Variant size (tiny/small/medium):** Larger variants are gradually more accurate but slower. Whether the difference is noticeable depends on your specific cameras and scenes. A good rule of thumb is to use the largest model your hardware can run without skipping detections, which you can monitor on the <NavPath path="Health and Metrics > Cameras" /> page in the UI. Better accuracy only helps if your detector keeps up with the detection load across all cameras.
**Acceptable inference time depends on your hardware.** Inference time alone does not tell the whole story, because different hardware has different capacity. A GPU can run multiple instances of the same model concurrently, so an inference time around 30ms can still keep up with several cameras. A Google Coral runs only a single instance of the model, so it needs a much lower inference time (around 10ms) to keep up.
The best detection accuracy comes from a model trained on images that look like what Frigate actually sees: security camera footage cropped to regions of interest. You can train or fine-tune your own model on images like this and run it as a custom model (see the per-detector sections below), but [Frigate+](/plus) makes this much easier by handling the training for you on images submitted from your own cameras. For YOLOv9, the `s` (small) variant at `320x320` resolution is a good place to start.
Frigate provides a number of builtin detector types. By default, Frigate will use a single CPU detector. Other detectors may require additional configuration as described below. Each of a model's devices runs in a dedicated process, and they pull from a common queue of detection requests from the cameras assigned to that model.
The Edge TPU detector type runs TensorFlow Lite models utilizing the Google Coral delegate for hardware acceleration. To use it, prefix a model's device with `edgetpu`.
The Edge TPU device can be specified using the `"device"` attribute according to the [Documentation for the TensorFlow Lite Python API](https://coral.ai/docs/edgetpu/multiple-edgetpu/#using-the-tensorflow-lite-python-api). If not set, the delegate will use the first device it finds.
Navigate to <NavPath path="Settings > System > Detection models" /> and select **Coral EdgeTPU (USB)** from the **Hardware** dropdown and check each Coral the model should run on.
Navigate to <NavPath path="Settings > System > Detection models" /> and select **Coral EdgeTPU (PCIe)** from the **Hardware** dropdown and check each Coral the model should run on.
Navigate to <NavPath path="Settings > System > Detection models" /> and select **Coral EdgeTPU (USB)** from the **Hardware** dropdown. USB and PCIe Corals are listed as separate hardware, so mixing the two on one model has to be done in YAML.
This detector is available for use with the Hailo-8, Hailo-8L and Hailo-8R AI Acceleration Modules. The integration identifies which of them is attached and selects the matching default model if no custom model is specified.
If no custom model is provided, the Hailo detector downloads a default model from the Hailo Model Zoo on first startup. Once cached, the model works fully offline. See [Network Requirements](/frigate/network_requirements#hardware-specific-detector-models) for details.
The HailoRT runtime is not part of the Frigate image. It is downloaded and installed into `/config/.local` the first time a Hailo detector is configured, verified against pinned checksums, and updated automatically when a Frigate release pins a new version. If the container has no internet access, see [Detector runtimes](/frigate/network_requirements#detector-runtimes) for how to provide the files yourself.
If both are provided, the detector will first check for the model at the given local path. If the file is not found, it will download the model from the specified URL. The model file is cached under `/config/model_cache/hailo`.
Hailo supports all models in the Hailo Model Zoo that include HailoRT post-processing. You're welcome to choose any of these pre-configured models for your implementation.
> The config.path parameter can accept either a local file path or a URL ending with .hef. When provided, the detector will first check if the path is a local file path. If the file exists locally, it will use it directly. If the file is not found locally or if a URL was provided, it will attempt to download the model from the specified URL.
The OpenVINO detector type runs an OpenVINO IR model on AMD and Intel CPUs, Intel GPUs and Intel NPUs. To use it, prefix a model's device with `openvino`.
The OpenVINO device to be used is specified using the `"device"` attribute according to the naming conventions in the [Device Documentation](https://docs.openvino.ai/2025/openvino-workflow/running-inference/inference-devices-and-modes.html). The most common devices are `CPU`, `GPU`, or `NPU`.
OpenVINO is supported on 6th Gen Intel platforms (Skylake) and newer. It will also run on AMD CPUs despite having no official support for it. A supported Intel platform is required to use the `GPU` or `NPU` device with OpenVINO. For detailed system requirements, see [OpenVINO System Requirements](https://docs.openvino.ai/2025/about-openvino/release-notes-openvino/system-requirements.html)
**NPU + GPU Systems:** If you have both NPU and GPU available (Intel Core Ultra processors), use NPU for object detection and GPU for enrichments (semantic search, face recognition, etc.) for best performance and compatibility.
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:
The NPU device must be passed into the container by adding `/dev/accel:/dev/accel` to the `devices` section of your compose file. Frigate grants the runtime user access to the device automatically; see [hardware device access](/configuration/non_root#hardware-device-access) if you manage device permissions yourself.
The NPU firmware is loaded by the host kernel and is not part of the Frigate image. Everything else the NPU needs is bundled in the container, so host NPU libraries should never be mounted in.
Frigate bundles a specific version of Intel's [linux-npu-driver](https://github.com/intel/linux-npu-driver/releases), and the host firmware must come from that release or a newer one. Firmware older than the bundled driver may fail with `MAPPED_INFERENCE_VERSION is NOT compatible with the ELF`, where `Expected` is the version the firmware supports and `received` is the version the bundled compiler produced. Distributions often package older firmware than the driver Frigate ships, so check the build date on the host with `sudo dmesg | grep -i vpu` and update it there if needed.
Intel NPUs cannot be used under Home Assistant OS, which does not include the NPU firmware.
The Apple Silicon detector client is being reworked. Its extra options no longer have a place in the config, so only the endpoint carried in the device string is honored right now, and `request_timeout_ms` and `linger_ms` are ignored. Anything else is dropped when your config is migrated.
The NPU in Apple Silicon can't be accessed from within a container, so the [Apple Silicon detector client](https://github.com/frigate-nvr/apple-silicon-detector) must first be setup. It is recommended to use the Frigate docker image with `-standard-arm64` suffix, for example `ghcr.io/blakeblackshear/frigate:stable-standard-arm64`.
Support for AMD GPUs is provided using the [ONNX detector](#onnx). In order to utilize the AMD GPU for object detection use a frigate docker image with `-rocm` suffix, for example `ghcr.io/blakeblackshear/frigate:stable-rocm`.
| gfx1103 (RDNA3 integrated) | Radeon 780M / 760M in Ryzen 7040, 8040, 8000G, and 200 series processors |
| gfx1150, gfx1151, gfx1152, gfx1153 (RDNA3.5) | Integrated GPUs in Ryzen AI 300 / 400 series and Ryzen AI Max processors |
| gfx1200, gfx1201 (RDNA4) | Radeon RX 9060 / 9070 series, Radeon AI PRO R9700 |
Other RDNA2 GPUs (gfx1031 through gfx1036), such as the Radeon RX 6700 / 6600 / 6500 series and the Radeon 680M / 660M integrated GPUs, run on the gfx1030 support by [overriding the chipset](#docker-settings-for-overriding-the-gpu-chipset) to `10.3.0`.
Older GPUs, including RDNA1 (Radeon RX 5000 series) and Vega based dedicated and integrated GPUs, are not supported.
ROCm needs access to the `/dev/kfd` and `/dev/dri` devices. When docker or frigate is not run under root then also `video` (and possibly `render` and `ssl/_ssl`) groups should be added.
When running docker directly the following flags should be added for device access:
```bash
$ docker run --device=/dev/kfd --device=/dev/dri \
For reference on recommended settings see [running ROCm/pytorch in Docker](https://rocm.docs.amd.com/projects/install-on-linux/en/develop/how-to/3rd-party/pytorch-install.html#using-docker-with-pytorch-pre-installed).
### Docker settings for overriding the GPU chipset
A GPU whose chipset isn't in the [supported GPUs](#supported-gpus-rocm) list can run on the support for a closely related chipset by setting the `HSA_OVERRIDE_GFX_VERSION` environment variable. RDNA2 GPUs other than gfx1030 (gfx1031 through gfx1036) need this, using `10.3.0`:
GPUs in the supported list don't need an override. An override that was set for the Radeon 780M / 760M (gfx1103) on an earlier version of Frigate is no longer needed and can be removed.
You can't tell the chipset from the AMD brand name, so check it in the container:
- first make sure that the ROCm environment is running properly by running `/opt/rocm/bin/rocminfo` in the Frigate container. It should list both the CPU and the GPU with their properties
- find the chipset version you have (gfxNNN) from the output of `rocminfo` (see below)
- if the chipset isn't in the supported list, set `HSA_OVERRIDE_GFX_VERSION` to the closest supported chipset, written as major.minor.stepping (for example gfx1030 is `10.3.0`)
- if things are not working check the Frigate docker logs
The AMD GPU kernel is known problematic especially when converting models to mxr format. The recommended approach is:
1. Disable object detection in the config.
2. Startup Frigate with the onnx detector configured, the main object detection model will be converted to mxr format and cached in the config directory.
3. Once this is finished as indicated by the logs, enable object detection in the UI and confirm that it is working correctly.
ONNX is an open format for building machine learning models, Frigate supports running ONNX models on CPU, OpenVINO, ROCm, TensorRT, and a Mac's Neural Engine. On startup Frigate will automatically try to use a GPU if one is available.
- The Neural Engine will automatically be detected and used with the ONNX detector when Frigate runs under lighter with its Neural Engine device. See [Apple Neural Engine (lighter)](#apple-neural-engine-lighter).
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:
### Apple Neural Engine (lighter) {#apple-neural-engine-lighter}
[lighter](https://github.com/fieldwork-ai/lighter) is an open-source container runtime for macOS. A container started with its `lighter.sh/ane` device gets an ONNX Runtime execution provider that runs models on the Mac's Neural Engine, and the ONNX detector uses it automatically, with the same models and configuration as on any other hardware. It works on M1 and newer Macs with lighter 0.9.2 or newer.
Give the Frigate container the Neural Engine device. With Docker Compose:
Or with `docker run`, add `--device lighter.sh/ane=all`. Frigate then reports the Neural Engine under **Settings > System > Detection models**, and the ONNX detector's model loads on it. lighter can also decode camera streams on the Mac's media engine; see [Video Decoding](hardware_acceleration_video.md#apple-silicon-mac-lighter).
The CPU detector type runs a TensorFlow Lite model utilizing the CPU without hardware acceleration. It is recommended to use a hardware accelerated detector type instead for better performance. To use it, set a model's device to `cpu`.
The CPU detector is not recommended for general use. If you do not have GPU or Edge TPU hardware, using the [OpenVINO Detector](#openvino-detector) in CPU mode is often more efficient than using the CPU detector.
:::
The number of threads used by the interpreter can be specified using the `"num_threads"` attribute, and defaults to `3.`
A TensorFlow Lite model is provided in the container at `/cpu_model.tflite` and is used by this detector type by default. To provide your own model, bind mount the file into the container and provide the path with the model's `path`.
This detector is available for use with the MemryX MX3 accelerator M.2 module. Frigate supports the MX3 on compatible hardware platforms, providing efficient and high-performance object detection.
The MemryX SDK is not part of the Frigate image. It is downloaded and installed into `/config/.local` the first time a MemryX detector is configured, verified against pinned checksums, and updated automatically when a Frigate release pins a new version. If the container has no internet access, see [Detector runtimes](/frigate/network_requirements#detector-runtimes) for how to provide the files yourself.
To use your own custom model, first compile it into a [.dfp](https://developer.memryx.com/2p1/specs/files.html#dataflow-program) file, which is the format used by MemryX.
> **Note:** It is recommended to compile the model on the host machine, or on another separate machine, rather than inside the Frigate Docker container. Installing the compiler inside Docker may conflict with container packages. It is recommended to create a Python virtual environment and install the compiler there.
For detailed instructions on compiling models, refer to the [MemryX Compiler](https://developer.memryx.com/2p1/tools/neural_compiler.html#usage) docs and [Tutorials](https://developer.memryx.com/2p1/tutorials/tutorials.html).
1. Package your compiled model into a `.zip` file.
2. The `.zip` file must contain the compiled `.dfp` file.
3. Depending on the model, the compiler may also generate a cropped post-processing network. If present, it will be named with the suffix `_post.onnx`.
This detector is available for use with the DEEPX NPU, both the DX-M1 M.2 module and the DX-M1M on the Sixfab AI HAT+ for the Raspberry Pi 5. The configuration below applies unchanged to either form factor. DEEPX NPU support in Frigate is developed and maintained by [Sixfab](https://sixfab.com).
See the [installation docs](../frigate/installation.md#deepx-npu) for information on installing the DEEPX kernel driver and runtime on the host and passing the NPU through to the container.
To run a model on a DEEPX NPU, list a `deepx` device on that model.
:::info
The DX-RT Python bindings are not part of the Frigate image. They are downloaded and installed into `/config/.local` the first time a DEEPX device is configured, verified against pinned checksums, and updated automatically when a Frigate release pins a new version. If the container has no internet access, see [Detector runtimes](/frigate/network_requirements#detector-runtimes) for how to provide the files yourself.
Frigate does not bundle a model for this detector. Models must be compiled to DEEPX's `.dxnn` format. Two model types are supported:
- `yolo-generic` for YOLO object detection models, the recommended default. The detector reads the model's output layout from the compiled file, so anchor-based, anchor-free and NMS-in-head models all work with the same configuration, as do models compiled with DEEPX's Post-Processing Unit (PPU) support.
- `yolox` for YOLOX models compiled without PPU support, whose raw head needs Frigate's YOLOX decoder. A YOLOX model compiled with PPU support works under either `yolox` or `yolo-generic`.
The quickest way to get one is the [DEEPX ModelZoo](https://developer.deepx.ai/modelzoo), which publishes pre-compiled `.dxnn` files for a range of YOLO object detection models. Download the `.dxnn`, bind-mount it into the container, and point the model's `path` at it. Alternatively, compile your own model with the DX-COM compiler. The recommended starting point is `yolox-s_640x640_ppu.dxnn`, the fastest ModelZoo model measured through Frigate:
For PPU models, use a `.dxnn` compiled with DX-COM 2.4.0 or later. Frigate reads the PPU head layout the compiler writes into the file and refuses to load a PPU model without it.
`model_type` must be set to `yolo-generic` or `yolox` to match the model; `yolo-generic` is the recommended default unless the model is a raw YOLOX export. Frigate defaults it to `ssd`, which this detector does not support, so the detector refuses to start on a model that leaves it unset.
`width` and `height` must match the resolution the model was compiled for. Quantization parameters are baked into the `.dxnn` file at compile time, so no normalization is applied on the host and Frigate's default `input_tensor`, `input_pixel_format`, and `input_dtype` values do not need to be overridden.
A DEEPX device is `PCIe:<index>`, as reported on the detector settings page. The NPU daemon multiplexes across processes, so the same device may be listed more than once to run additional inference processes against it:
```yaml
models:
- devices:
- deepx:PCIe:0
- deepx:PCIe:0
```
#### Label maps
The object detection models in the DEEPX ModelZoo are trained on the standard 80-class COCO label set, so `labelmap_path` must be set to `/labelmap/coco-80.txt`. Frigate's default label map uses an extended 91-class COCO scheme, and leaving it in place will cause detections to be reported as the wrong object type. For `yolo-generic` models the label map is also what the detector uses to tell the output layout, so a label map with the wrong number of classes is reported as an error at startup.
Nvidia Jetson devices may be used for object detection using the TensorRT libraries. Due to the size of the additional libraries, this detector is only provided in images with the `-tensorrt-jp6` tag suffix, e.g. `ghcr.io/blakeblackshear/frigate:stable-tensorrt-jp6`. This detector is designed to work with Yolo models for object detection.
### Generate Models
The model used for TensorRT must be preprocessed on the same hardware platform that they will run on. This means that each user must run additional setup to generate a model file for the TensorRT library. A script is included that will build several common models.
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, 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.
If your GPU does not support FP16 operations, you can pass the environment variable `USE_FP16=False` to disable it.
Specific models can be selected by passing an environment variable to the `docker run` command or in your `docker-compose.yml` file. Use the form `-e YOLO_MODELS=yolov4-416,yolov4-tiny-416` to select one or more model names. The models available are shown below.
<details>
<summary>Available Models</summary>
```
yolov3-288
yolov3-416
yolov3-608
yolov3-spp-288
yolov3-spp-416
yolov3-spp-608
yolov3-tiny-288
yolov3-tiny-416
yolov4-288
yolov4-416
yolov4-608
yolov4-csp-256
yolov4-csp-512
yolov4-p5-448
yolov4-p5-896
yolov4-tiny-288
yolov4-tiny-416
yolov4x-mish-320
yolov4x-mish-640
yolov7-tiny-288
yolov7-tiny-416
yolov7-640
yolov7-416
yolov7-320
yolov7x-640
yolov7x-320
```
</details>
An example `docker-compose.yml` fragment that converts the `yolov4-608` and `yolov7x-640` models would look something like this:
```yml
frigate:
environment:
- YOLO_MODELS=yolov7-320,yolov7x-640
- USE_FP16=false
```
### Configuration Parameters
The TensorRT detector can be selected by specifying `tensorrt` as the model type. The GPU will need to be passed through to the docker container using the same methods described in the [Hardware Acceleration](hardware_acceleration_video.md#nvidia-gpus) section. If you pass through multiple GPUs, you can select which GPU is used for a detector with the `device` configuration parameter. The `device` parameter is an integer value of the GPU index, as shown by `nvidia-smi` within the container.
The TensorRT detector uses `.trt` model files that are located in `/config/model_cache/tensorrt` by default. These model path and dimensions used will depend on which model you have generated.
Use the config below to work with generated TRT models:
Hardware accelerated object detection is supported on the following SoCs:
- SL1680
This implementation uses the [Synaptics model conversion](https://synaptics-synap.github.io/doc/v/latest/docs/manual/introduction.html#offline-model-conversion), version v3.1.0.
This implementation is based on sdk `v1.5.0`.
See the [installation docs](../frigate/installation.md#synaptics) for information on configuring the SL-series NPU hardware.
If no custom model is provided, the RKNN detector downloads a default model from GitHub on first startup. Once cached, the model works fully offline. See [Network Requirements](/frigate/network_requirements#hardware-specific-detector-models) for details.
When using many cameras one detector may not be enough to keep up. Multiple detectors can be defined assuming NPU resources are available. An example configuration would be:
This `config.yml` shows all relevant options to configure the detector and explains them. All values shown are the default values (except for two). Lines that are required at least to use the detector are labeled as required, all other lines are optional.
The inference time was determined on a rk3588 with 3 NPU cores.
- All models are automatically downloaded and stored in the folder `config/model_cache/rknn_cache`. After upgrading Frigate, you should remove older models to free up space.
- You can also provide your own `.rknn` model. You should not save your own models in the `rknn_cache` folder, store them directly in the `model_cache` folder or another subfolder. To convert a model to `.rknn` format see the `rknn-toolkit2` (requires a x86 machine). Note, that there is only post-processing for the supported models.
- Place one or more models in onnx format in the directory `config/model_cache/rknn_cache/onnx` on your docker host (this might require `sudo` privileges).
- Save the configuration file under `config/conv2rknn.yaml` (see below for details).
- Run `docker exec <frigate_container_id> python3 /opt/conv2rknn.py`. If the conversion was successful, the rknn models will be placed in `config/model_cache/rknn_cache`.
This is an example configuration file that you need to adjust to your specific onnx model:
- `soc`: A list of all SoCs you want to build the rknn model for. If you don't specify this parameter, the script tries to find out your SoC and builds the rknn model for this one.
- `quantization`: true: 8 bit integer (i8) quantization, false: 16 bit float (fp16). Default: false.
- `output_name`: The output name of the model. The following variables are available:
- `soc`: the SoC this model was build for (e.g. "rk3588")
- `tk_version`: Version of `rknn-toolkit2` (e.g. "2.3.0")
- **example**: Specifying `output_name = "frigate-{quant}-{input_basename}-{soc}-v{tk_version}"` could result in a model called `frigate-i8-my_model-rk3588-v2.3.0.rknn`.
- `config`: Configuration passed to `rknn-toolkit2` for model conversion. For an explanation of all available parameters have a look at section "2.2. Model configuration" of [this manual](https://github.com/MarcA711/rknn-toolkit2/releases/download/v2.3.2/03_Rockchip_RKNPU_API_Reference_RKNN_Toolkit2_V2.3.2_EN.pdf).
The AXEngine detector downloads its default model from HuggingFace on first startup. Once cached, the model works fully offline. See [Network Requirements](/frigate/network_requirements#hardware-specific-detector-models) for details.
The AXEngine python package is not part of the Frigate image. It is downloaded and installed into `/config/.local` the first time an AXEngine detector is configured, verified against a pinned checksum, and updated automatically when a Frigate release pins a new version. If the container has no internet access, see [Detector runtimes](/frigate/network_requirements#detector-runtimes) for how to provide the file yourself.
Hardware accelerated object detection is supported on the Axelera Metis AI Accelerator, available in M.2 and PCIe card formats. Inference runs through Axelera's low-level runtime (`axelera.runtime`), which executes compiled YOLO-family models with a flat per-frame latency.
See the [installation docs](../frigate/installation.md#axelera-metis) for information on configuring the Metis hardware.
:::info
The Axelera runtime is not part of the Frigate image. It is downloaded and installed into `/config/.local` the first time an Axelera detector is configured, verified against pinned checksums, and updated automatically when a Frigate release pins a new version. If the container has no internet access, see [Detector runtimes](/frigate/network_requirements#detector-runtimes) for how to provide the files yourself.
Models must be compiled for the Metis NPU with the Axelera Voyager SDK before Frigate can run them. The compiled output is a directory containing `1/model.json`, `1/manifest.json`, and the baked `1/postprocess_graph.onnx`.
Compile the model on a machine with the Voyager SDK installed (compilation is not supported inside the Frigate container), for example with the SDK's `deploy.py` against a model-zoo yaml. Then zip the compiled output and point the model's `path` at the zip (bind-mounted, or at an http(s) URL); the archive extracts fresh on every start.
