UltraFace

UltraFace is a lightweight face detector designed for edge computing devices.

The official code can be found here.

Weights are available in the releases section of the repository:

• version-RFB-640.onnx: data/models/face_detector_ultraface/version-RFB-640.onnx

• version-RFB-320.onnx: data/models/face_detector_ultraface/version-RFB-320.onnx

Performance on WIDER Face Val dataset

Backbone	Input size (W, H)	AP - easy	AP - medium	AP - hard	Inference time (s/img) - CPU	Inference time (s/img) - GPU
version-RFB-640	(640, 480)	85.5%	82.2%	57.9%	0.0675	0.0191
version-RFB-320	(320, 240)	78.7%	69.8%	43.8%	0.0251	0.0085

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

Usage example

import cv2
from toolbox.Models.face_detector_ultraface import FaceDetector

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

detector = FaceDetector(
    model_path="../../../data/models/face_detector_ultraface/version-RFB-640.onnx",
    input_size=(640, 480),
    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(0.128,0.552,0.358,0.780) 0.99999964
# > BoundingBox(0.580,0.542,0.806,0.840) 0.99998796
# > BoundingBox(0.600,0.075,0.854,0.343) 0.999987
# > BoundingBox(0.108,0.082,0.290,0.312) 0.99997556

Project configuration YAML example:

face_detector:
    model_name: face_detector_ultraface
    params:
      model_path: ../../../data/models/face_detector_ultraface/version-RFB-320.onnx
      input_size: [320, 240]
      use_cuda: True