Β«YOLO5FaceΒ» reproduced the paper "YOLO5Face: Why Reinventing a Face Detector"
ONLINE DEMOοΌFACE AND LANDMARKS DETECT
ARCH | GFLOPs | Easy | Medium | Hard | |
---|---|---|---|---|---|
zjykzj/YOLO5Face (This) | yolov5s-face | 15.2 | 94.69 | 93.00 | 84.73 |
deepcam-cn/yolov5-face(Official) | yolov5s-face | / | 94.33 | 92.61 | 83.15 |
zjykzj/YOLO5Face (This) | shufflenetv2-face | 1.5 | 90.27 | 87.39 | 73.60 |
deepcam-cn/yolov5-face(Official) | shufflenetv2-face | / | 90.76 | 88.12 | 73.82 |
zjykzj/YOLO5Face (This) | yolov5x-v7.0 | 204 | 95.79 | 94.53 | 87.63 |
zjykzj/YOLO5Face (This) | yolov5s-v7.0 | 15.8 | 94.84 | 93.28 | 84.67 |
zjykzj/YOLO5Face (This) | yolov5n-v7.0 | 4.2 | 93.25 | 91.11 | 80.33 |
- Table of Contents
- Newsπππ
- Background
- Installation
- Usage
- Maintainers
- Thanks
- Contributing
- License
Version | Release Date | Major Updates |
---|---|---|
v1.1.1 | 2024/08/17 | Update and add Gradio demo. |
v1.1.0 | 2024/07/21 | Supports additional models including shufflenetv2-face/yolov5x-v7.0/yolov5n-v7.0. |
v1.0.0 | 2024/07/14 | Adds keypoint detection, enabling face + keypoint detection. |
v0.1.0 | 2024/06/29 | Trains a face detector based on yolov5-v7.0 and the WIDERFACE dataset. |
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 ultralytics/yolov5, 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.txt
Or use docker container
$ docker run -it --runtime nvidia --gpus=all --shm-size=16g -v /etc/localtime:/etc/localtime -v $(pwd):/workdir --workdir=/workdir --name yolo5face ultralytics/yolov5:v7.0
- Download WIDERFACE dataset: Baidu Drive(08p0)
# yolov5s_v7.0
$ python3 widerface_train.py --data widerface.yaml --weights "" --cfg models/yolo5face/cfgs/yolov5s_v7_0.yaml --hyp models/yolo5face/hyps/hyp.scratch-low.yaml --img 800 --epoch 300 --device 0
# yolov5s-face
$ python3 widerface_train.py --data widerface.yaml --weights "" --cfg models/yolo5face/cfgs/yolov5s_face.yaml --hyp models/yolo5face/hyps/hyp.scratch.yaml --img 800 --epoch 300 --device 0
$ python widerface_detect.py --weights ./runs/exp4-yolov5s_v7_0-i800-e300.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_v7_0 summary: 157 layers, 7039792 parameters, 0 gradients, 15.8 GFLOPs
...
Speed: 0.4ms pre-process, 8.8ms inference, 0.8ms NMS per image at shape (1, 3, 640, 640)
Results saved to runs/detect/exp7
0 labels saved to runs/detect/exp7/labels
$ cd widerface_evaluate/
$ python3 evaluation.py -p ../runs/detect/exp7/labels/ -g ./ground_truth/
Reading Predictions : 100%|βββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββ| 61/61 [00:00<00:00, 62.18it/s]
Processing easy: 100%|ββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββ| 61/61 [00:20<00:00, 2.94it/s]
Processing medium: 100%|ββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββ| 61/61 [00:20<00:00, 2.98it/s]
Processing hard: 100%|ββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββ| 61/61 [00:20<00:00, 2.97it/s]
==================== Results ====================
Easy Val AP: 0.9483604102331251
Medium Val AP: 0.9328484206418773
Hard Val AP: 0.8467345083774318
=================================================
$ python detect_face_and_landmarks.py --weights ./runs/exp4-yolov5s_v7_0-i800-e300.pt --source assets/selfie.jpg --imgsz 2048 --conf-thres 0.25 --iou-thres 0.45 --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 Β© 2024 zjykzj