Awesome
yolov7-pose
Implementation of "YOLOv7: Trainable bag-of-freebies sets new state-of-the-art for real-time object detectors"
Pose estimation implimentation is based on YOLO-Pose.
在yoloV7-pose基础上添加了任意关键点数量 + 多类别分类代码。
这里是以4个关键点进行举例，其中添加了左右翻转数据增强，
点的交换是：point1和point2 point3和point4 point5和point6 依次类推。设置了20个点的交换，可以取前n个，n为偶数。
Dataset preparison
# data.txt 含义分别是： cls x y w h point1x point1y point2x point2y point3x point3y point4x point4y ...
# 类别 目标中心点x 目标中心点y 目标宽w 目标高h 目标点1x坐标 目标点1y坐标 目标点2x坐标 目标点2y坐标 目标点3x坐标 目标点3y坐标 目标点4x坐标 目标点4y坐标 依次类推
0 0.5739299610894941 0.1724137931034483 0.3715953307392996 0.29064039408866993 0.38910505836575876 0.08374384236453201 0.7587548638132295 0.029556650246305417 0.7607003891050583 0.2660098522167488 0.39299610894941633 0.32019704433497537
2 0.5739299610894941 0.1724137931034483 0.3715953307392996 0.29064039408866993 0.38910505836575876 0.08374384236453201 0.7587548638132295 0.029556650246305417 0.7607003891050583 0.2660098522167488 0.39299610894941633 0.32019704433497537
0 0.5739299610894941 0.1724137931034483 0.3715953307392996 0.29064039408866993 0.38910505836575876 0.08374384236453201 0.7587548638132295 0.029556650246305417 0.7607003891050583 0.2660098522167488 0.39299610894941633 0.32019704433497537
写一个train.txt 和 val.txt文件
# train.txt
./train/images/-nfs-阿拉伯车牌字符-沙特阿拉伯卡口车牌-2-沙特阿拉伯卡口车牌-2-image1837.jpeg
./train/images/-nfs-车牌字符-埃及车牌-埃及车牌截图-2021-04-30 11-11-52屏幕截图.png
./train/images/-nfs-车牌字符-埃及车牌-埃及车牌截图-2021-04-30 13-57-27屏幕截图.png
./train/images/-nfs-车牌字符-埃及车牌-埃及车牌截图-2021-04-30 10-19-54屏幕截图.png
./train/images/-nfs-阿拉伯车牌字符-外国车牌现场_20210519_1-外国车牌现场_20210519_1-e0d92b0990a1249388bc77bdfa8e43ed.jpg
./train/images/-nfs-车牌字符-埃及车牌-埃及车牌截图-2021-04-30 13-51-28屏幕截图.png
./train/images/-nfs-车牌字符-约旦车牌-videoplayback-videoplayback_13_1460.jpg
./train/images/-nfs-车牌字符-埃及车牌-埃及车牌截图-2021-04-30 13-56-51屏幕截图.png
./train/images/-nfs-车牌字符-埃及车牌-埃及车牌截图-2021-04-30 10-27-50屏幕截图.png
Training
百度网盘：yolov7-w6-person.pt 提取码: 9nlk
python -m torch.distributed.launch --nproc_per_node 8 --master_port 9527 train_Ncla_nPoint.py --data data/coco_kpts.yaml --cfg cfg/yolov7-w6-pose.yaml --weights weights/yolov7-w6-person.pt --batch-size 128 --img 640 --kpt-label --sync-bn --device 0,1,2,3,4,5,6,7 --name yolov7-w6-pose --hyp data/hyp.pose.yaml
Deploy
TensorRT:https://github.com/nanmi/yolov7-pose
Testing
python test_Ncla.py --data data/coco_kpts.yaml --img 640 --conf 0.5 --iou 0.25 --weights yolov7-w6-pose.pt --kpt-label