CV / RetinaFace

RetinaFace MobileNet-v1

RetinaFace (MobileNet-v1 backbone) detects faces plus 5 keypoints (eyes, nose, mouth corners) on Hailo-8. The HEF exposes 9 raw heads; anchors, SSD decode, NMS and landmark decode run on the host CPU.

1 ダウンロード

サイズ

6.3 MB

メモリ

4GB+

精度

Hailo HEF / HailoRT 4.23.x

使用しているデバイスを選択してください。セットアップガイドとドキュメントがそれに応じて更新されます。

プラットフォーム

推論エンジン

はじめる

デプロイ

sudo docker run --rm \
  --name cm5-hailo8-retinaface \
  --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/retinaface_mobilenet_v1:latest \
  python web_detection.py --model_path model/retinaface_mobilenet_v1.hef --video_path video/test.mp4

REST API

REST API を使用して推論を実行します。以下のコマンドをコピーしてください。

Curl

curl -X POST "http://<Board_IP>:8000/api/models/retinaface_mobilenet_v1/predict" \
  -F "file=@test.jpg"

Python

import requests

response = requests.post(
    "http://<Board_IP>:8000/api/models/retinaface_mobilenet_v1/predict",
    files={"file": open("test.jpg", "rb")},
    timeout=30,
)
print(response.json())

モデル詳細

RetinaFace MobileNet-v1 on reComputer R Series (CM5 + Hailo-8)

RetinaFace detects faces — plus 5 keypoints (eyes, nose tip, mouth corners) per face — on Hailo-8 through HailoRT. The HEF exposes 9 raw heads and the decode runs on the host CPU.

Model Info

Property	Value
Architecture	RetinaFace, MobileNet-v1 backbone (biubug6/Pytorch_Retinaface)
Task	Face detection
Input	736x1280x3 BGR (normalize_in_net mean [123,117,104])
Output	9 heads: 3 scales x {bbox, conf, landmark}
Priors	38,640 anchors (2/cell, feature maps 92x160/46x80/23x40)
Classes	1 (face) + 5 landmarks per detection
Parameters	3.49M
Operations	25.14G
mAP (WIDER FACE)	81.3% hardware reference
HEF	Hailo Model Zoo v2.19.0, Hailo-8

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

Run With Demo Video

bash

sudo docker run --rm \
  --name cm5-hailo8-retinaface \
  --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/retinaface_mobilenet_v1:latest \
  python web_detection.py --model_path model/retinaface_mobilenet_v1.hef --video_path video/test.mp4

Open http://<Board_IP>:8000 to view the web preview (face boxes + green keypoints overlaid on the MJPEG stream).

REST API

bash

curl -X POST "http://<Board_IP>:8000/api/models/retinaface_mobilenet_v1/predict" \
  -F "file=@test.jpg"

Endpoint	Method	Purpose
`/`	GET	Web preview UI
`/api/models/retinaface_mobilenet_v1/predict`	POST	Faces + landmarks (JSON)
`/api/video_feed`	GET	MJPEG preview stream
`/api/config`	GET / POST	Confidence / NMS IOU thresholds

Implementation Notes

CPU post-processing (meta_arch=retinaface): no on-chip NMS. Anchor generation, SSD decode (variances 10/5), softmax confidence, greedy NMS and landmark decode are a numpy port of the Model Zoo's face_detection_postprocessing.py.
Head layout per the official reshapes: 8-channel head = bbox (2 anchors x 4 coords), 4-channel head = confidence (2 anchors x {background, face}), 20-channel head = landmarks (2 anchors x 10 coords).
The alls script applies input_conversion(bgr_to_rgb) in the HEF — the app feeds raw uint8 BGR frames.
Preprocessing: aspect-ratio-preserving resize + bottom/right pad (color 0) to 736x1280.

Development Notes

Source module: src/rpi5_hailo8_retinaface_mobilenet_v1/
Dockerfile: docker/hailo8/retinaface_mobilenet_v1.dockerfile
Container: ghcr.io/seeed-projects/recomputer-hailo8-cv/retinaface_mobilenet_v1:latest
Family: RetinaFace (variant: mobilenet_v1)

入力と出力

Input: image, video, or USB camera frame. Output: face boxes with confidences, 5 keypoints per face (eyes, nose, mouth corners), and an annotated MJPEG preview.