RetinaFace

RetinaFace is a single-stage face detector implemented in PyTorch based on the original paper “RetinaFace: Single-stage Dense Face Localisation in the Wild”.

The official code can be found here.

Weights are available in the releases section of the repository:

Resnet50: data/models/face_detector_retinaface/mobilenet0.25_Final.pth
Mobilenet0.25: data/models/face_detector_retinaface/Resnet50_Final.pth

Performance on WIDER Face Val dataset

Backbone	AP - easy	AP - medium	AP - hard	Inference time (s/img) - CPU	Inference time (s/img) - GPU
Resnet50	95.48%	94.04%	84.43%	3.617	0.136
Mobilenet0.25	90.70%	88.16%	73.82%	0.253	0.120

^{CPU: Intel(R) Xeon(R) Silver 4116 || GPU: Quadro RTX 8000}

Usage example

import cv2
from toolbox.Models.face_detector_retinaface import FaceDetector

img = cv2.imread("data/samples/images/faces/celeb/ben_mad_min_jer.png")

detector = FaceDetector(
    weights_path="data/models/face_detector_retinaface/mobilenet0.25_Final.pth", 
    model_name="mobile0.25",
    use_cuda=False
)
instances = detector.predict(img)

print(instances[0].fields)
# > ['bounding_box', 'confidence']

for instance in instances:
    print(instance.bounding_box, instance.confidence)
# > BoundingBox(0.130,0.565,0.360,0.800) 0.9999852
# > BoundingBox(0.604,0.077,0.838,0.350) 0.9998944
# > BoundingBox(0.580,0.565,0.800,0.853) 0.9998043
# > BoundingBox(0.104,0.073,0.300,0.313) 0.99966705

Project configuration YAML example:

face_detector:
    model_name: face_detector_retinaface
    params:
      weights_path: ../../../data/models/face_detector_retinaface/Resnet50_Final.pth
      model_name: resnet50
      confidence_threshold: 0.7
      landmarks: False
      nms_threshold: 0.4
      max_input_size: 512
      use_cuda: True

Note:

The model may struggle to detect large faces since it was trained with anchors ranging from 16 to 512 pixels. To address this issue, use the parameter max_input_size to scale down larger images (e.g. 512). It can also be used to speed up inference.