SSD MobileNet V2 on reComputer R Series (CM5 + Hailo-8)

SSD MobileNet V2 is the classic TensorFlow SSD detector with a MobileNet V2 backbone. It detects 80 COCO object classes on Hailo-8. Same I/O as V1 but lighter and slightly more accurate. NMS runs on-chip (Hailo HPP, meta_arch=ssd), so the host only parses the post-NMS tensor.

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

Model Info

Hardware and Host Setup

sudo apt update
sudo apt install hailort hailort-pcie-driver python3-hailort
sudo reboot
hailortcli --version
hailortcli fw-control identify
ls -l /dev/hailo0

Run With Demo Video

sudo docker run --rm \
  --name cm5-hailo8-ssd-mobilenet-v2 \
  --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/ssd_mobilenet_v2:latest \
  python web_detection.py --model_path model/ssd_mobilenet_v2.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-ssd-mobilenet-v2 \
  --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/ssd_mobilenet_v2:latest \
  python web_detection.py --model_path model/ssd_mobilenet_v2.hef --camera_id 0

REST API

POST http://<Board_IP>:8000/api/models/ssd_mobilenet_v2/predict

curl -X POST "http://<Board_IP>:8000/api/models/ssd_mobilenet_v2/predict" \
  -F "file=@test.jpg"

Implementation Notes

• Identical I/O contract and post-processing to the V1 build; only the backbone differs (MobileNet V2 → lighter, slightly more accurate).
• The HEF runs NMS on-chip; the app only parses the post-NMS tensor (tf_postproc_nms), so nms_thresh is ignored (API parity only).
• normalize_in_net mean=127.5/std=127.5 (classic SSD normalization); the app feeds raw uint8 RGB pixels after letterboxing — no manual normalization.
• HailoRT returns the NMS vstream as a ragged per-class list (NMS-by-score); the parser handles ragged/object/dense layouts. First inference logs the raw type/shape so it can be verified on hardware.
• Class mapping: cls_id (0..89) → COCO category ID cls_id+1 (labels_offset=1); 10 unused COCO IDs are "N/A" and not drawn.
• Defaults: confidence 0.25, IOU 0.45 (IOU no effect — NMS on-chip). Eval uses 0.3 / 0.6.

Development Notes

• Source module: src/rpi5_hailo8_ssd_mobilenet_v2/
• Dockerfile: docker/hailo8/ssd_mobilenet_v2.dockerfile
• Container: ghcr.io/seeed-projects/recomputer-hailo8-cv/ssd_mobilenet_v2:latest
• Family: ssd (variants ssd_mobilenet_v1 / ssd_mobilenet_v2); this is the V2 build.