ArcFace MobileFaceNet on reComputer R Series (CM5 + Hailo-8)

ArcFace with a MobileFaceNet backbone extracts 512-dim face embeddings on Hailo-8 through HailoRT. The embedding is used for face verification and identification; face comparison itself is done by the client with cosine similarity.

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

# Both candidates must be 4.23.x before installation.
apt-cache policy hailort hailort-pcie-driver

sudo apt install hailort=4.23.0 hailort-pcie-driver=4.23.0
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. Do not continue if apt-cache policy selects HailoRT 5.x; that runtime targets Hailo-10H and does not match this Hailo-8 container.

Run With Demo Video

sudo docker run --rm \
  --name cm5-hailo8-arcface \
  --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/arcface_mobilefacenet:latest \
  python web_detection.py --model_path model/arcface_mobilefacenet.hef --video_path video/test.mp4

Open http://<Board_IP>:8000 to view the web preview. Each frame is annotated with the embedding dimension, and the REST API returns the full 512-dim vector.

USB Camera Mode

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

REST API

Prediction endpoint:

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

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

{
  "success": true,
  "source": "uploaded image",
  "embedding": [0.023, -0.156, 0.089, "..."],
  "dimension": 512,
  "image": { "width": 1280, "height": 720 }
}

Face verification is not exposed as a separate endpoint: compare two returned embeddings with cosine similarity on the client side (L2-normalize both vectors first).

Implementation Notes

• The HEF contains its own normalization layer, so the application letterboxes frames to 112x112, converts BGR to RGB, and feeds raw uint8 pixels.
• The embedding is read from the arcface_mobilefacenet/fc1 output when the vstream name is exposed; otherwise the first output tensor is used. The first inference logs every vstream name and shape so the mapping can be verified on hardware.
• Empty or non-finite embeddings are rejected instead of being returned as a valid result.
• --class_path is accepted for CLI compatibility but has no effect: ArcFace has no class label table.

Development Notes

• Source module: src/rpi5_hailo8_arcface_mobilefacenet/
• Dockerfile: docker/hailo8/arcface_mobilefacenet.dockerfile
• Container: ghcr.io/seeed-projects/recomputer-hailo8-cv/arcface_mobilefacenet:latest
• HEF: model/arcface_mobilefacenet.hef (4,100,573 bytes)
• The module uses the HailoRT 4.23 VStreams API and a Hailo-8-specific HEF.