EfficientDet-Lite2 on reComputer R Series (CM5 + Hailo-8)

EfficientDet-Lite2 performs COCO 80-class object detection on Hailo-8 through HailoRT. NMS and sigmoid run on-chip (Hailo HPP), so the host only parses the post-NMS tensor. This is the Lite2 variant of the efficientdet family — largest backbone and 448x448 input for the highest accuracy of the three (lite0/lite1/ lite2).

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 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 this image is 4.23.x.

Run With Demo Video

sudo docker run --rm \
  --name cm5-hailo8-efficientdet-lite2 \
  --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/efficientdet_lite2:latest \
  python web_detection.py --model_path model/efficientdet_lite2.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-efficientdet-lite2 \
  --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/efficientdet_lite2:latest \
  python web_detection.py --model_path model/efficientdet_lite2.hef --camera_id 0

REST API

Prediction endpoint:

POST http://<Board_IP>:8000/api/models/efficientdet_lite2/predict

Example image request:

curl -X POST "http://<Board_IP>:8000/api/models/efficientdet_lite2/predict" \
  -F "file=@test.jpg"

Implementation Notes

• Identical input/output contract and post-processing to the Lite0/Lite1 builds; only the backbone and input size differ (448x448, 5.93M params → highest accuracy of the three).
• The HEF runs NMS on-chip (device_pre_post_layers: nms=true, sigmoid=true, hpp=true); the app only parses the post-NMS tensor (per the official tf_postproc_nms), so nms_thresh is ignored (API parity only).
• Post-NMS rows are [ymin, xmin, ymax, xmax, score], normalized to [0,1] of the 448x448 letterboxed input; the app scales to pixels and un-letterboxes.
• normalize_in_net with mean=127/std=128 + padding_color=127: the app letterboxes with gray (127) padding and feeds raw uint8 RGB pixels — no manual normalization.
• HailoRT returns the NMS vstream as a ragged per-class list (NMS-by-score); the parser handles that plus object/dense layouts. First inference logs the raw type/shape so it can be verified on hardware.
• Class mapping: cls_id (0..88) → COCO category ID cls_id+1 (labels_offset=1); the 10 unused COCO IDs are "N/A" and not drawn. If labels look shifted, re-derive from the first-inference log.
• Defaults: confidence 0.25, IOU 0.45 (IOU has no effect — NMS on-chip). Model Zoo eval uses 0.001 / 0.5 — lower the confidence slider to inspect more.

Development Notes

• Source module: src/rpi5_hailo8_efficientdet_lite2/
• Dockerfile: docker/hailo8/efficientdet_lite2.dockerfile
• Container: ghcr.io/seeed-projects/recomputer-hailo8-cv/efficientdet_lite2:latest
• The module uses the HailoRT 4.23 VStreams API and a Hailo-8-specific HEF.
• Family: efficientdet (variants lite0 / lite1 / lite2); this is the lite2 build.