YOLOv10

YOLOv10 on reComputer R Series (CM5 + Hailo-10H)

YOLOv10n / YOLOv10s / YOLOv10b / YOLOv10x perform COCO 80-class object detection on Hailo-10H through HailoRT. The compiled HEF runs NMS on-chip (HPP) with the one2one head, so the host parses the post-NMS result and never decodes boxes or runs NMS itself.

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

YOLOv10n:

sudo docker run --rm \
  --name cm5-hailo10h-yolov10n \
  --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/yolov10n:latest \
  python web_detection.py --model_path model/yolov10n.hef --video_path video/test.mp4

YOLOv10s:

sudo docker run --rm \
  --name cm5-hailo10h-yolov10s \
  --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/yolov10s:latest \
  python web_detection.py --model_path model/yolov10s.hef --video_path video/test.mp4

YOLOv10b:

sudo docker run --rm \
  --name cm5-hailo10h-yolov10b \
  --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/yolov10b:latest \
  python web_detection.py --model_path model/yolov10b.hef --video_path video/test.mp4

YOLOv10x:

sudo docker run --rm \
  --name cm5-hailo10h-yolov10x \
  --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/yolov10x:latest \
  python web_detection.py --model_path model/yolov10x.hef --video_path video/test.mp4

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

USB Camera Mode

sudo docker run --rm \
  --name cm5-hailo10h-yolov10n \
  --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/yolov10n:latest \
  python web_detection.py --model_path model/yolov10n.hef --camera_id 0

Swap yolov10n for yolov10s, yolov10b or yolov10x to use another variant.

HEF Files

All four variants are Hailo Model Zoo v5.4.0 Hailo-10H builds of the same THU-MIG YOLOv10 family.

Source: https://hailo-model-zoo.s3.eu-west-2.amazonaws.com/ModelZoo/Compiled/v5.4.0/hailo10h/yolov10n.hef (yolov10s.hef, yolov10b.hef, yolov10x.hef in the same directory).

REST API

POST http://<Board_IP>:8000/api/models/yolov10n/predict

curl -X POST "http://<Board_IP>:8000/api/models/yolov10n/predict" \
  -F "file=@test.jpg"

Implementation Notes

• The Model Zoo base/yolov10.yaml uses hpp=true and device_pre_post_layers.nms=true with meta_arch=yolov8, applies sigmoid on the three one2one class convolutions and documents info.output_shape: 80x5x100 (80 classes, at most 100 boxes per class).
• The parser nodes are the one2one head (/model.23/one2one_cv2.{0,1,2}/Conv box, /model.23/one2one_cv3.{0,1,2}/Conv class): one prediction per grid cell, no anchors. Nothing is decoded and no NMS runs on the host.
• HailoRT 5.1.1 returns the raw NMS-by-class buffer as a flat 1-D FLOAT32 array in compact layout: per class one integer-valued detection count followed by only that many [ymin, xmin, ymax, xmax, score] rows, normalized to [0, 1]. The parser stops at the first implausible class header, so the worst-case padding tail is not read as data; ragged NMS-by-score and dense (1,C,5,D) / (1,C,D,5) layouts stay supported as well.
• The IOU slider has no effect (NMS is on-chip); the confidence slider filters the parsed detections.
• The HEF bakes normalize_in_net (mean 0 / std 255) and padding_color=114: the app letterboxes with gray (114) and feeds raw uint8 RGB pixels — no manual normalization.
• Class IDs 0..79 index the standard COCO class list directly.
• First-inference diagnostics are disabled in production (_OUTPUT_LOGGED = True); set it to False in web_detection.py to dump the raw output type, shape, value range and per-class rows when validating a new HEF.

Development Notes

• Source modules: src/hailo10h_yolov10{n,s,b,x}/ in the reComputer-Hailo10H-CV repository
• Containers: ghcr.io/seeed-projects/recomputer-hailo10h-cv/yolov10{n,s,b,x}:latest
• On Hailo-8 (R20 demo) only n/b/x are published; the Hailo-10H builds cover all four variants.
• Model licence: AGPL-3.0 (upstream THU-MIG/yolov10).