YOLOv11-obb on reComputer R Series (CM5 + Hailo-10H)

YOLOv11n / YOLOv11s / YOLOv11m / YOLOv11x-OBB perform oriented object detection (rotated boxes, DOTA v1, 15 classes) on Hailo-10H through HailoRT. The compiled HEF has no on-chip NMS: it exposes nine raw heads and the host decodes them and runs the rotated probabilistic-IoU NMS — the same decode the Hailo Model Zoo uses for its published DOTA mAP.

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

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

The expected HailoRT version for these images is 5.1.1.

Run With Demo Video

YOLOv11n_obb:

sudo docker run --rm \
  --name cm5-hailo10h-yolov11n_obb \
  --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/yolov11n_obb:latest \
  python web_detection.py --model_path model/yolov11n_obb.hef --video_path video/test.mp4

YOLOv11s_obb:

sudo docker run --rm \
  --name cm5-hailo10h-yolov11s_obb \
  --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/yolov11s_obb:latest \
  python web_detection.py --model_path model/yolov11s_obb.hef --video_path video/test.mp4

YOLOv11m_obb:

sudo docker run --rm \
  --name cm5-hailo10h-yolov11m_obb \
  --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/yolov11m_obb:latest \
  python web_detection.py --model_path model/yolov11m_obb.hef --video_path video/test.mp4

YOLOv11x_obb:

sudo docker run --rm \
  --name cm5-hailo10h-yolov11x_obb \
  --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/yolov11x_obb:latest \
  python web_detection.py --model_path model/yolov11x_obb.hef --video_path video/test.mp4

Open http://<Board_IP>:8000 to view the web preview.

The bundled demo clip is a short aerial (drone) sample, so the DOTA-trained model detects vehicles and other aerial objects directly and the rotated boxes show up in the preview; drone-dataset classes that are not part of DOTA (e.g. pedestrians) are not reported.

USB Camera Mode

sudo docker run --rm \
  --name cm5-hailo10h-yolov11n_obb \
  --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/yolov11n_obb:latest \
  python web_detection.py --model_path model/yolov11n_obb.hef --camera_id 0

Swap yolov11n_obb for yolov11s_obb, yolov11m_obb or yolov11x_obb to use another variant.

HEF Files

All four variants are Hailo Model Zoo v5.4.0 Hailo-10H builds of the same THU-MIG / ultralytics YOLOv11-OBB family (DOTA v1). mAP and FPS are the Model Zoo's published Hailo-10H numbers.

Source: https://hailo-model-zoo.s3.eu-west-2.amazonaws.com/ModelZoo/Compiled/v5.4.0/hailo10h/yolov11n_obb.hef (yolov11s_obb.hef, yolov11m_obb.hef, yolov11x_obb.hef in the same directory).

REST API

POST http://<Board_IP>:8000/api/models/yolov11n_obb/predict

curl -X POST "http://<Board_IP>:8000/api/models/yolov11n_obb/predict" \
  -F "file=@aerial.jpg" -F "conf=0.25" -F "iou=0.7"

Each prediction returns the class, the confidence and the rotated box:

{
  "predictions": [
    {
      "class": "ship",
      "confidence": 0.94,
      "obb": {"x_center": 450.0, "y_center": 320.0, "width": 180.0, "height": 420.0, "angle": 35.2}
    }
  ]
}

angle is in degrees (counter-clockwise from +x, range [-45, 135)).

Implementation Notes

• The Model Zoo base/yolo_obb.yaml sets device_pre_post_layers.sigmoid: false, meta_arch: yobb and score_threshold: 0.01, with no hpp and no on-chip NMS; anchors uses scale_factors [0.5, 0.5], regression_length 15 (16 DFL bins) and strides [8, 16, 32], and evaluation is DOTA v1 with 15 classes.
• Head layout: nine raw tensors, identified by channel count (64-channel DFL box / 15-channel class / 1-channel angle) and paired per stride (128x128, 64x64, 32x32).
• Angle: theta = (sigmoid(a) - 0.25) * pi (range [-pi/4, 3pi/4)), exactly as in the Model Zoo's YoloOBBPostProc and ultralytics' OBB head.
• Boxes: the DFL softmax expectation over the 16 bins gives the l/t/r/b distances in stride units; w = l + r, h = t + b, the centre is the anchor plus the rotated half-axis offset, and the distances are scaled by the stride.
• Selection: best class per anchor above the confidence slider, then probabilistic-IoU Fast-NMS with a per-class offset (max_wh 7680, at most 30000 candidates in and 300 boxes out); the confidence and IOU sliders are both live because the NMS runs on the host.
• The HEF bakes normalize_in_net (mean 0 / std 255) and padding_color=114: the app letterboxes to 1024x1024 with gray (114) and feeds raw uint8 RGB pixels — no manual normalization.
• Class IDs 0..14 index the DOTA v1 list (plane, ship, storage-tank, baseball-diamond, ...) as in the Model Zoo's dotav1_names.json.
• First-inference diagnostics are disabled in production (_DET_OUTPUT_LOGGED = True); set it to False in web_detection.py to dump the head layout, value ranges and the decoded angle span when validating a new HEF.

Development Notes

• Source modules: src/hailo10h_yolov11{n,s,m,x}_obb/ in the reComputer-Hailo10H-CV repository
• Containers: ghcr.io/seeed-projects/recomputer-hailo10h-cv/yolov11{n,s,m,x}_obb:latest
• Hailo-8 has no official OBB build (the Model Zoo lists Hailo-8/Hailo-8L as NA for oriented object detection); yolov11l_obb targets Hailo-15L only.
• Model licence: AGPL-3.0 (upstream ultralytics/ultralytics).