This repo gave multi-task keypoint detect code based yolov8. The landmarks or keypoints with different classes and numbers can be detected at the same time.
⭐ This code has been completely released and the documentation will be improved gradually.⭐
-This repo is almost the raw yolov8. And the following repo or article are referred:
###Based aboving work, we made a little upgrade:
⭐ The pose-keypoints or landmarks detection based yolov8 was achieved, which is anchor free and owns decoupled head.
⭐ The muticlass with different points number detection was achieved, and these parameters can be set conveniently by .yaml.
⭐ ⭐ The most important, we introduced a new val.py for point accuracy measurement, such as:
Class Images Instances Box(P R mAP50 mAP50-95 AVG_distance MAX_distance interval(0,1] interval(1,2] interval(2,3] interval(3,4] interval(4,5] interval(5,6] interval(6,7] interval(7,8] interval(8,9] interval(9,10]interval(10,+∞] kpt_p kpt_r kpt_p_x kpt_r_x kpt_p_y kpt_r_y): 100%|██████████| 51/51 [01:54<00:00, 2.24s/it] all 407 74821 0.929 0.84 0.915 0.591 1.18 223 0.311 0.312 0.157 0.0781 0.0442 0.0274 0.0181 0.0123 0.00837 0.00625 0.0253 0.956 0.955 0.994 0.992 0.969 0.967
###The sample model config file is v8/yolov8s.yaml. And the point number config file is datasets/test.yaml.