wmcnally / kapao

KAPAO is an efficient single-stage human pose estimation model that detects keypoints and poses as objects and fuses the detections to predict human poses.
GNU General Public License v3.0
750 stars 103 forks source link

How about support Halpe dataset? #41

Open lucasjinreal opened 2 years ago

lucasjinreal commented 2 years ago

Hi, kappa seems very promising, but I think it can be more powerful. Would u consider add Halpe dataset support? Here I am not mean add a whole body key points , but can u train a body25 key points model?

body25 can be much more useful then simple 17 key points in coco. It can provides foot motion as well and you simply don't need changing anything just dataset.

Halpe dataset link: (it just using coco and MPII images) https://github.com/Fang-Haoshu/Halpe-FullBody

wmcnally commented 2 years ago

It is on my TO-DO list to train on COCO-WholeBody

lucasjinreal commented 2 years ago

@wmcnally that's look nice. wholebody is nice, but kind of slow in terms of useful scenarios. if a body25 or body26 version avaiable we can make kapao more useful.

wmcnally commented 2 years ago

COCO-WholeBody has basically the same keypoints as Halpe-FullBody but way more examples. I agree that having the feet and hands would be useful.

lofyol commented 2 years ago

Thank you for your work. Have you finished the model training of body-25?