xingyizhou
mix ctdet and multi_pose
I want to integrate detection and key point, I think this method is feasible, My goal is to detect the key points of people when they are detected while detecting the target
for this reason, I changed the annotation of coco key data set, I added other categories except people to the coco key point annotation, and then attached 0 to the key point coordinates and 0 to the number of key points of these categories.
I changed the relevant corresponding code and changed the num_classes to 3
elif opt.task == 'multi_pose': # assert opt.dataset in ['coco_hp'] opt.flip_idx = dataset.flip_idx opt.heads = {'hm': opt.num_classes, 'wh': 2, 'hps': 34} if opt.reg_offset: opt.heads.update({'reg': 2}) if opt.hm_hp: opt.heads.update({'hm_hp': 17}) if opt.reg_hp_offset: opt.heads.update({'hp_offset': 2})
The model results I got:
My problem is that except for the category of people, other categories cannot obtain accurate results
