CV / YOLACT

YOLACT-RegNetX-1.6GF

YOLACT (UC Davis) with a RegNetX-1.6GF backbone predicts instance masks with 49,104 anchors, Fast NMS and prototype-mask assembly (proto @ coeffs) on Hailo-8. Higher accuracy than the 800MF variant; post-processing runs on the host CPU since the HEF exposes 16 raw heads.

2 téléchargements

Taille

36 MB

Mémoire

4GB+

Précision

Hailo HEF / HailoRT 4.23.x

Choisissez l'appareil que vous utilisez. Le guide de configuration et la documentation seront mis à jour en conséquence.

Plateforme

Moteur d'inférence

Pour commencer

Déployer

sudo docker run --rm \
  --name cm5-hailo8-yolact-1-6gf \
  --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/yolact_regnetx_1_6gf:latest \
  python web_detection.py --model_path model/yolact_regnetx_1_6gf.hef --video_path video/test.mp4

API REST

Utilisez l'API REST pour exécuter l'inférence. Copiez les commandes ci-dessous.

Curl

curl -X POST "http://<Board_IP>:8000/api/models/yolact_regnetx_1_6gf/predict" \
  -F "file=@test.jpg"

Python

import requests

response = requests.post(
    "http://<Board_IP>:8000/api/models/yolact_regnetx_1_6gf/predict",
    files={"file": open("test.jpg", "rb")},
    timeout=30,
)
print(response.json())

Détails du modèle

YOLACT-RegNetX-1.6GF on reComputer R Series (CM5 + Hailo-8)

YOLACT (UC Davis) performs real-time instance segmentation — per-object masks on top of COCO 80-class boxes — on Hailo-8 through HailoRT. The HEF exposes 16 raw heads and all decode runs on the host CPU. The 1.6GF variant is the larger, more accurate member of the yolact family.

Model Info

Property	Value
Architecture	YOLACT (You Only Look At CoefficienTs), RegNetX-1.6GF backbone
Task	Instance segmentation
Input	512x512x3 RGB (normalize_in_net ImageNet RGB mean/std)
Output	16 heads: proto (128x128x32) + 5 scales x {bbox, mask coeffs, conf}
Priors	49,104 anchors (9/cell, feature maps 64/32/16/8/4)
Classes	80 (COCO)
Parameters	30.09M
Operations	125.34G
mAP (COCO, full precision)	27.6% 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-yolact-1-6gf \
  --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/yolact_regnetx_1_6gf:latest \
  python web_detection.py --model_path model/yolact_regnetx_1_6gf.hef --video_path video/test.mp4

Open http://<Board_IP>:8000 to view the web preview (colored instance masks overlaid on the MJPEG stream).

REST API

bash

curl -X POST "http://<Board_IP>:8000/api/models/yolact_regnetx_1_6gf/predict" \
  -F "file=@test.jpg"

Endpoint	Method	Purpose
`/`	GET	Web preview UI
`/api/models/yolact_regnetx_1_6gf/predict`	POST	Boxes, confidences, mask areas (JSON)
`/api/video_feed`	GET	MJPEG preview stream with masks
`/api/config`	GET / POST	Confidence / NMS IOU thresholds

Implementation Notes

CPU post-processing (meta_arch=yolact): no on-chip NMS. The Detect pipeline (49,104 anchors → SSD decode → per-instance argmax → Fast NMS, top-200/class, max 100 detections) is a numpy port of the Model Zoo's instance_segmentation_postprocessing.py.
Candidate masks: sigmoid(proto @ coeffs^T) cropped to each detection box; visualized with mask_thresh=0.5 (YAML mask_threshold).
normalize_in_net ImageNet RGB; the app feeds raw uint8 RGB.
Preprocessing: plain bilinear resize to 512x512 (no letterbox, no pad).

Development Notes

Source module: src/rpi5_hailo8_yolact_regnetx_1_6gf/
Dockerfile: docker/hailo8/yolact_regnetx_1_6gf.dockerfile
Container: ghcr.io/seeed-projects/recomputer-hailo8-cv/yolact_regnetx_1_6gf:latest
Family: YOLACT (variants: regnetx_800mf / regnetx_1.6gf)

Entrées et sorties

Input: image, video, or USB camera frame. Output: COCO 80-class boxes with confidences, per-object mask pixel areas, and an MJPEG preview with colored instance masks.