CenterNet (resnet_v1_50) on reComputer R Series (CM5 + Hailo-8)

CenterNet with a ResNet-50 backbone performs COCO 80-class object detection on Hailo-8 through HailoRT, using the on-chip max_finder post-processing variant. It is the higher-accuracy sibling of the resnet_v1_18 build: same input/output contract, deeper backbone.

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

Model Info

The accuracy value is Hailo's Model Zoo reference. It is not a benchmark measured on CM5.

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 hailo-all
sudo reboot

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

The expected HailoRT version for this image is 4.23.x.

Run With Demo Video

sudo docker run --rm \
  --name cm5-hailo8-centernet-r50 \
  --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/centernet_resnet_v1_50_postprocess:latest \
  python web_detection.py --model_path model/centernet_resnet_v1_50_postprocess.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-centernet-r50 \
  --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/centernet_resnet_v1_50_postprocess:latest \
  python web_detection.py --model_path model/centernet_resnet_v1_50_postprocess.hef --camera_id 0

REST API

Prediction endpoint:

POST http://<Board_IP>:8000/api/models/centernet_resnet_v1_50_postprocess/predict

Example image request:

curl -X POST "http://<Board_IP>:8000/api/models/centernet_resnet_v1_50_postprocess/predict" \
  -F "file=@test.jpg"

Implementation Notes

• Identical input/output contract and post-processing to the resnet_v1_18 build; only the backbone differs (ResNet-50 → more accurate, heavier).
• The HEF was compiled with normalize_in_net=true (ImageNet mean/std, RGB), so the app letterboxes to 512x512, converts BGR to RGB, and feeds raw uint8 pixels — no manual normalization.
• The heatmap head is already sparse: the on-chip max_finder + score threshold (0.2) keep only local maxima above threshold. The decoder takes those peaks, reads wh/reg at each peak, and reconstructs boxes with the official Hailo Model Zoo formula (stride 4, 128x128 feature map -> 512x512 input), then un-letterboxes back to the original frame.
• The two 2-channel heads (wh, reg) are mapped by vstream name when the conv number is exposed (conv3 -> wh, conv5 -> reg), otherwise by network output order. First inference prints every vstream name/shape and the resolved mapping so it can be verified on hardware.

Development Notes

• Source module: src/rpi5_hailo8_centernet_resnet_v1_50_postprocess/
• Dockerfile: docker/hailo8/centernet_resnet_v1_50_postprocess.dockerfile
• Container: ghcr.io/seeed-projects/recomputer-hailo8-cv/centernet_resnet_v1_50_postprocess:latest
• The module uses the HailoRT 4.23 VStreams API and a Hailo-8-specific HEF.