YOLO26 on reComputer R Series (CM5 + Hailo-8)

YOLO26n / YOLO26s / YOLO26m perform COCO 80-class object detection on Hailo-8 through HailoRT. YOLO26 is a one2one head (one prediction per grid cell, no anchors, no NMS): the host decodes the raw split heads and selects with the two-stage top-k, and a compile that ships the on-chip HPP NMS result is parsed as well — the layout in use is logged on the first inference.

This page targets reComputer R Series (CM5 + Hailo-8) with a PCIe Hailo-8 accelerator.

Hardware and Host Setup

Install the Hailo-8 packages on the host and confirm that the driver and device are ready before starting the container:

sudo apt update
sudo apt install hailort hailort-pcie-driver python3-hailort
sudo reboot

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

The expected HailoRT version for these images is 4.23.x.

Run With Demo Video

YOLO26n:

sudo docker run --rm \
  --name cm5-hailo8-yolo26n \
  --privileged \
  --net=host \
  -e PYTHONUNBUFFERED=1 \
  --device /dev/hailo0:/dev/hailo0 \
  -v /usr/lib/libhailort.so.4.23.0:/usr/lib/libhailort.so.4.23.0:ro \
  -v /usr/lib/libhailort.so:/usr/lib/libhailort.so:ro \
  ghcr.io/seeed-projects/recomputer-hailo8-cv/yolo26n:latest \
  python web_detection.py --model_path model/yolo26n.hef --video_path video/test.mp4

YOLO26s:

sudo docker run --rm \
  --name cm5-hailo8-yolo26s \
  --privileged \
  --net=host \
  -e PYTHONUNBUFFERED=1 \
  --device /dev/hailo0:/dev/hailo0 \
  -v /usr/lib/libhailort.so.4.23.0:/usr/lib/libhailort.so.4.23.0:ro \
  -v /usr/lib/libhailort.so:/usr/lib/libhailort.so:ro \
  ghcr.io/seeed-projects/recomputer-hailo8-cv/yolo26s:latest \
  python web_detection.py --model_path model/yolo26s.hef --video_path video/test.mp4

YOLO26m:

sudo docker run --rm \
  --name cm5-hailo8-yolo26m \
  --privileged \
  --net=host \
  -e PYTHONUNBUFFERED=1 \
  --device /dev/hailo0:/dev/hailo0 \
  -v /usr/lib/libhailort.so.4.23.0:/usr/lib/libhailort.so.4.23.0:ro \
  -v /usr/lib/libhailort.so:/usr/lib/libhailort.so:ro \
  ghcr.io/seeed-projects/recomputer-hailo8-cv/yolo26m:latest \
  python web_detection.py --model_path model/yolo26m.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-hailo8-yolo26n \
  --privileged \
  --net=host \
  -e PYTHONUNBUFFERED=1 \
  --device /dev/hailo0:/dev/hailo0 \
  --device /dev/video0:/dev/video0 \
  -v /usr/lib/libhailort.so.4.23.0:/usr/lib/libhailort.so.4.23.0:ro \
  -v /usr/lib/libhailort.so:/usr/lib/libhailort.so:ro \
  ghcr.io/seeed-projects/recomputer-hailo8-cv/yolo26n:latest \
  python web_detection.py --model_path model/yolo26n.hef --camera_id 0

Swap yolo26n for yolo26s or yolo26m to use another variant.

HEF Files

All three variants are Hailo Model Zoo v2.19.0 Hailo-8 builds of the same Ultralytics YOLO26 family.

Source: https://hailo-model-zoo.s3.eu-west-2.amazonaws.com/ModelZoo/Compiled/v2.19.0/hailo8/yolo26n.hef (yolo26s.hef, yolo26m.hef in the same directory).

REST API

POST http://<Board_IP>:8000/api/models/yolo26n/predict

curl -X POST "http://<Board_IP>:8000/api/models/yolo26n/predict" \
  -F "file=@test.jpg"

Implementation Notes

• The Model Zoo base/yolo26.yaml uses nms=false, sigmoid=false, hpp=false and meta_arch=yolo26, with post_nms_topk=300 inherited from base/yolov8.yaml; info.output_shape documents 3 strides x (4 box, 80 class scores).
• One2one selection: no NMS runs at all. The host applies sigmoid, takes the top-k anchors by max class score, then the top-k (anchor, class) pairs — multiple classes per anchor are allowed — and filters by the confidence slider.
• Box decode: the 4-channel head gives the distances from the cell centre to the left/top and right/bottom edges in stride units (regression_length=1); 64-channel DFL compiles are decoded through the softmax expectation.
• If a compile instead returns the on-chip HPP NMS result (as the yolo26-seg HEFs do) the post-NMS tensor is parsed (compact per-class buffer, dense Cx5xD / CxDx5, ragged NMS-by-score list) and no host selection runs.
• 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.
• The first inference logs the layout and the box/score ranges ([YOLO26] layout=..., outputs=[...]) so it can be confirmed on hardware.

Development Notes

• Source modules: src/hailo10h_yolo26{n,s,m}/ (Hailo-10H) or src/rpi5_hailo8_yolo26{n,s,m}/ (Hailo-8) in the matching model repository
• Containers: ghcr.io/seeed-projects/recomputer-hailo8-cv/yolo26{n,s,m}:latest
• The detection family ships n/s/m; YOLO26l / YOLO26x have no Hailo HEF yet.
• Model licence: AGPL-3.0 (upstream ultralytics/ultralytics).