YOLOv8-pose

YOLOv8 Pose on reComputer R Series (CM5 + Hailo-10H)

YOLOv8 Pose estimates human keypoints on Hailo-10H through HailoRT. The model detects people, returns one box per person together with 17 COCO keypoints, and serves a browser preview, single-image API and offline video analysis.

This page targets reComputer R Series (CM5 + Hailo-10H) with a PCIe Hailo-10H accelerator.

Model Info

Hailo Model Zoo publishes s and m for the Hailo-10H build. Each size ships as its own image (yolov8s_pose and yolov8m_pose), so the pulled image identifies the model. s is the default here.

Hardware and Host Setup

Install the Hailo-10H packages on the host and confirm that the driver and device are ready before starting the container:

sudo apt update
sudo apt install hailo-h10-all -y
sudo reboot

# After reboot
hailortcli --version
hailortcli fw-control identify
ls -l /dev/hailo0

Run With Demo Video

sudo docker run --rm \
  --name cm5-hailo10h-yolov8s-pose \
  --privileged \
  --net=host \
  -e PYTHONUNBUFFERED=1 \
  --device /dev/hailo0:/dev/hailo0 \
  -v /usr/lib/libhailort.so.5.1.1:/usr/lib/libhailort.so.5.1.1:ro \
  -v /usr/lib/libhailort.so:/usr/lib/libhailort.so:ro \
  ghcr.io/seeed-projects/recomputer-hailo10h-cv/yolov8s_pose:latest \
  python web_detection.py --model_path model/yolov8s_pose.hef --video_path video/test.mp4

Open http://<Board_IP>:8000 to view the preview: person boxes with the COCO skeleton drawn on top.

Larger Model

sudo docker run --rm \
  --name cm5-hailo10h-yolov8m-pose \
  --privileged \
  --net=host \
  -e PYTHONUNBUFFERED=1 \
  --device /dev/hailo0:/dev/hailo0 \
  -v /usr/lib/libhailort.so.5.1.1:/usr/lib/libhailort.so.5.1.1:ro \
  -v /usr/lib/libhailort.so:/usr/lib/libhailort.so:ro \
  ghcr.io/seeed-projects/recomputer-hailo10h-cv/yolov8m_pose:latest \
  python web_detection.py --model_path model/yolov8m_pose.hef --video_path video/test.mp4

USB Camera Mode

sudo docker run --rm \
  --name cm5-hailo10h-yolov8s-pose \
  --privileged \
  --net=host \
  -e PYTHONUNBUFFERED=1 \
  --device /dev/hailo0:/dev/hailo0 \
  --device /dev/video0:/dev/video0 \
  -v /usr/lib/libhailort.so.5.1.1:/usr/lib/libhailort.so.5.1.1:ro \
  -v /usr/lib/libhailort.so:/usr/lib/libhailort.so:ro \
  ghcr.io/seeed-projects/recomputer-hailo10h-cv/yolov8s_pose:latest \
  python web_detection.py --model_path model/yolov8s_pose.hef --camera_id 0

REST API

Prediction endpoint:

POST http://<Board_IP>:8000/api/models/yolov8_pose/predict

curl -X POST "http://<Board_IP>:8000/api/models/yolov8_pose/predict" \
  -F "file=@test.jpg"

{
  "success": true,
  "source": "uploaded image",
  "predictions": [
    {
      "class": "person",
      "confidence": 0.91,
      "box": {"x1": 120, "y1": 80, "x2": 360, "y2": 520},
      "keypoints": [{"x": 180, "y": 120, "score": 0.87}]
    }
  ],
  "image": {"width": 1280, "height": 720}
}

Keypoints below the score threshold (0.30) are omitted from the response.

Implementation Notes

• The HEF exposes the raw pose heads, there is no on-chip NMS: nine tensors at three feature-map scales, each with bbox DFL (64 channels), score (1) and keypoints (51).
• Decoding follows the ultralytics / Model Zoo pose head: sigmoid scores, DFL softmax for the box distances, keypoints as (2 * raw + grid) * stride in (x, y, score) order; NMS runs host-side.
• Values already inside [0, 1] are kept as probabilities, otherwise sigmoid is applied, so the decoder matches either compiled form.

Development Notes

• Source module: src/hailo10h_yolov8s_pose/, src/hailo10h_yolov8m_pose/
• Dockerfile: docker/hailo10h/yolov8s_pose.dockerfile, docker/hailo10h/yolov8m_pose.dockerfile
• Container: ghcr.io/seeed-projects/recomputer-hailo10h-cv/yolov8s_pose:latest, ghcr.io/seeed-projects/recomputer-hailo10h-cv/yolov8m_pose:latest
• The module uses the HailoRT 5.1.1 InferModel API and a Hailo-10H-specific HEF.
• Unit tests (mocked HailoRT + pose post-process regression) run without a device: python -m unittest discover -s tests -v.