This repository contains script for inference RetinaFace with MobileNet encoder. The remake of the https://github.com/biubug6/Pytorch_Retinaface
pip install git+https://github.com/TsykunovDmitriy/retinaface_lightweight.pytorch
Make sure your CUDA version matches the latest PyTorch version. Otherwise, install the required version of PyTorch from here. To use tensorrt, install torch2trt from here
Inference
>>> import cv2
>>> from retinaface import RetinaDetector
>>> detector = RetinaDetector(
device="cpu", # or cuda or cuda:0 or number of GPU
score_thresh=0.5,
top_k=100,
nms_thresh=0.4,
use_trt=False,
)
>>> image = cv2.imread(path_to_image)
>>> bboxes, landmarks, scores = detector(image)
>>> print(bboxes, landmarks, scores)
(
[
[xmin, ymin, xmax, ymax]
],
[
[[x, y], # right eye
[x, y], # left eye
[x, y], # nose
[x, y], # right edge of the mouth
[x, y]] # left edge of the mouth
],
[score]
)Align
>>> aligned_face = detector.aligning(image, landmarks[person])Image/Video inference
python image_inference.py --image path_to_img.jpg --output path_to_save_result.jpg --device 0 --use_trt
python video_inference.py --video path_to_video.mp4 --output path_to_save_result.mp4 --device 0 --use_trt
