«YOLO5Face» reproduced the paper "YOLO5Face: Why Reinventing a Face Detector"
| ARCH | GFLOPs | Easy | Medium | Hard | |
|---|---|---|---|---|---|
| deepcam-cn/yolov5-face(Official) | yolov5s-face | / | 94.33 | 92.61 | 83.15 |
| zjykzj/YOLO5Face(This) | yolov5s-face | 15.1 | 94.34 | 92.72 | 84.47 |
| zjykzj/YOLO5Face(This) | yolov5s | 15.8 | 95.21 | 93.42 | 84.03 |
- [2024/06/29]v0.1.0. Training Face Detectors Based on YOLOv5-v7.0 and WIDERFACE Datasets.
YOLO5Face is a very interesting work that further abstracts the task of face detection. By directly using a universal object detection algorithm, good face detection results can be achieved. Of course, it also achieves 5-point facial keypoint regression. Based on the ultralytics/yolov5 object detection framework, YOLO5Face can easily apply different models and training, such as using lightweight networks for real-time detection and using large networks for higher detection accuracy.
Note: the latest implementation of YOLO5Face in our warehouse is entirely based on ultralytics/yolov5 v7.0
pip3 install -r requirements.txtOr use docker container
docker run -it --runtime nvidia --gpus=all --shm-size=16g -v /etc/localtime:/etc/localtime -v $(pwd):/workdir --workdir=/workdir --name yolov2 ultralytics/yolov5:latestDownload the WIDERFACE dataset from http://shuoyang1213.me/WIDERFACE/, Then convert WIDERFACE dataset format.
python3 widerface2yolo.py ../datasets/widerface/WIDER_train/images ../datasets/widerface/wider_face_split/wider_face_train_bbx_gt.txt ../datasets/widerface/
python3 widerface2yolo.py ../datasets/widerface/WIDER_val/images ../datasets/widerface/wider_face_split/wider_face_val_bbx_gt.txt ../datasets/widerface/# YOLOv5s
python3 train.py --data widerface.yaml --weights yolov5s.pt --cfg yolov5s.yaml --imgsz 800 --epoch 250 --device 0
# YOLOv5s-face
python3 train.py --data widerface.yaml --weights '' --cfg models/yolo5face/cfgs/yolov5s_face.yaml --hyp models/yolo5face/hyp.scratch.yaml --img 800 --epoch 250 --device 0# python widerface_detect.py --weights ./runs/train/exp4-yolov5s-e250-img800.pt --source ../datasets/widerface/images/val/ --folder_pict ../datasets/widerface/wider_face_split/wider_face_val_bbx_gt.txt --conf-thres 0.001 --iou-thres 0.6 --save-txt --save-conf --device 0
...
YOLOv5s summary: 157 layers, 7012822 parameters, 0 gradients, 15.8 GFLOPs
...
Speed: 0.3ms pre-process, 9.0ms inference, 0.9ms NMS per image at shape (1, 3, 640, 640)
Results saved to runs/detect/exp5
0 labels saved to runs/detect/exp5/labels
# cd widerface_evaluate/
# python3 evaluation.py -p ../runs/detect/exp5/labels/ -g ./ground_truth/
Reading Predictions : 100%|█████████████████████████████████████████████████████████████████████████████████████████████████████████| 61/61 [00:00<00:00, 94.45it/s]
Processing easy: 100%|██████████████████████████████████████████████████████████████████████████████████████████████████████████████| 61/61 [00:19<00:00, 3.13it/s]
Processing medium: 100%|████████████████████████████████████████████████████████████████████████████████████████████████████████████| 61/61 [00:19<00:00, 3.12it/s]
Processing hard: 100%|██████████████████████████████████████████████████████████████████████████████████████████████████████████████| 61/61 [00:19<00:00, 3.13it/s]
==================== Results ====================
Easy Val AP: 0.9520941964576021
Medium Val AP: 0.9341770033021547
Hard Val AP: 0.8403303849682994
=================================================python detect.py --weights ./runs/exp4-yolov5s-e250-img800.pt --source assets/selfie.jpg --imgsz 2048 --conf-thres 0.25 --iou-thres 0.45 --max-det 3000 --hide-labels --hide-conf
- zhujian - Initial work - zjykzj
Anyone's participation is welcome! Open an issue or submit PRs.
Small note:
- Git submission specifications should be complied with Conventional Commits
- If versioned, please conform to the Semantic Versioning 2.0.0 specification
- If editing the README, please conform to the standard-readme specification.
Apache License 2.0 © 2024 zjykzj