| --- |
| license: apache-2.0 |
| pipeline_tag: object-detection |
| base_model: yolo26n |
| tags: |
| - arm |
| - arm-optimized |
| - edge-linux |
| - raspberry-pi-5 |
| - xnnpack |
| --- |
| |
| # Arm Runtime for YOLO26 on Raspberry Pi 5 |
|
|
| ## About |
|
|
| This runtime runs YOLO26n on Raspberry Pi 5 with Raspberry Pi OS Trixie. It is |
| a fixed-graph implementation that uses model-specific memory planning, custom |
| operators, and operator fusion to maximize throughput on the Raspberry Pi CPU. |
|
|
| The runtime supports the standard YOLO26n model and compatible fine-tuned |
| variants at square resolutions of 320, 480, 640, 800, 960, and 1280 pixels. |
|
|
| ## Performance |
|
|
| The following YOLO26n results were measured on Raspberry Pi 5 at 640x640, |
| batch 1, against the 128-image, 1,025-instance |
| [`coco-128-by-sa`](https://huggingface.co/datasets/Arm/coco-128-by-sa/tree/97d6ea181e9dd0784db880b705e426ad23d424e1) |
| published dataset revision. Inference time excludes |
| preprocessing and postprocessing. |
|
|
| | Runtime | Inference (ms/image) | mAP50-95 | |
| | --- | ---: | ---: | |
| | Arm Runtime (this project) | 36.60 | 0.4863 | |
| | NCNN | 67.20 | 0.4902 | |
| | MNN | 90.90 | 0.4753 | |
| | OpenVINO | 102.00 | 0.4812 | |
| | ONNX Runtime | 126.90 | 0.4812 | |
| | ExecuTorch | 136.90 | 0.4880 | |
|
|
| See the [performance baseline](benchmarks/reporting/baseline.md) for the Arm |
| Runtime measurement and host context. |
|
|
| ## Quick start |
|
|
| ### Prerequisites |
|
|
| On Raspberry Pi OS Trixie, install the required packages: |
|
|
| ```bash |
| sudo apt update |
| sudo apt install --yes --no-install-recommends \ |
| ca-certificates \ |
| curl \ |
| git \ |
| libgl1 \ |
| libglib2.0-0t64 \ |
| make \ |
| patch \ |
| python3 \ |
| python3-venv \ |
| tar \ |
| gzip |
| ``` |
|
|
| Clone the repository: |
|
|
| ```bash |
| git clone https://github.com/arm/Optimized-Runtimes.git |
| cd Optimized-Runtimes/yolo26-raspberry-pi-5 |
| ``` |
|
|
| ### Run an example |
|
|
| Run object detection on a sample video using the standard YOLO26n checkpoint: |
|
|
| ```bash |
| make example-video |
| ``` |
|
|
| For a fine-tuned checkpoint, your own video, or live camera input, see |
| [Usage](#usage). |
|
|
| ### Run the benchmark |
|
|
| Run the [Ultralytics benchmark harness](https://docs.ultralytics.com/guides/raspberry-pi/#raspberry-pi-5-yolo26-benchmarks) |
| using this runtime: |
|
|
| ```bash |
| make ultralytics-benchmark |
| ``` |
|
|
| See the [Ultralytics benchmark documentation](benchmarks/ultralytics/README.md) |
| for configuration options. |
|
|
| ## Usage |
|
|
| ### Acquiring a runtime |
|
|
| A prebuilt runtime library is available from this repository's releases. The |
| example and benchmark targets download the release matching the checked-out |
| project version when no runtime exists under `dist/runtime/`. |
|
|
| To build the runtime from source, install the native build prerequisites: |
|
|
| ```bash |
| sudo apt update |
| sudo apt install --yes --no-install-recommends \ |
| build-essential \ |
| ca-certificates \ |
| cmake \ |
| curl \ |
| git \ |
| patch \ |
| python3 \ |
| python3-venv \ |
| tar \ |
| gzip |
| ``` |
|
|
| Then run: |
|
|
| ```bash |
| make runtime |
| ``` |
|
|
| The resulting runtime is installed under `dist/runtime/`, where the examples |
| and benchmark consume it automatically. |
|
|
| ### Acquiring a model |
|
|
| The examples and benchmark require a YOLO26n checkpoint converted for this |
| runtime. If the default model is absent, their Make targets download the pinned |
| standard YOLO26n checkpoint and convert it automatically. |
|
|
| The benchmark normally uses the standard model because its classes match the |
| default COCO dataset. To use a compatible fine-tuned checkpoint with the video |
| or camera examples, convert it first: |
|
|
| ```bash |
| make model CHECKPOINT=/path/to/your-checkpoint.pt |
| ``` |
|
|
| The default conversion uses 640x640 input and writes the model under |
| `dist/models/`. Select another supported resolution with, for example: |
|
|
| ```bash |
| make model CHECKPOINT=/path/to/your-checkpoint.pt RESOLUTION=320 |
| ``` |
|
|
| ## Examples and benchmarks |
|
|
| The examples use the [Python API](src/runtime/api/python/arm_yolo26n/README.md). |
| Their Make targets prepare missing default runtime and model artifacts |
| automatically. |
|
|
| - [Object detection on prerecorded video](examples/video/README.md) |
| - [Object detection on live video from a Raspberry Pi Camera](examples/picamera2/README.md) |
| - [Ultralytics benchmark](benchmarks/ultralytics/README.md) |
|
|
| ## Other information |
|
|
| Run `make help` for other useful commands. |
|
|
| ### Project layout |
|
|
| - `src/converter/` - Converts compatible YOLO26n PyTorch checkpoints into the runtime model format. |
| - `src/runtime/` - Implements the fixed YOLO26n runtime. |
| - `src/runtime/patches/` - Contains project changes to XNNPACK and pthreadpool. |
| - `src/runtime/api/python/arm_yolo26n/` - Provides the Python API. |
| - `examples/` - Contains prerecorded-video and live-camera examples. |
| - `benchmarks/ultralytics/` - Integrates the runtime with the Ultralytics benchmark. |
|
|
