search

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
751 stars 103 forks source link

Does keypoints head share conv with detection head? #76

Open lucasjinreal opened 2 years ago

lucasjinreal commented 2 years ago

Does keypoints head share conv with detection head?

wmcnally commented 2 years ago

Yes!

wmcnally commented 2 years ago

Keypoints and poses are simply treated as different object classes.

lucasjinreal commented 2 years ago

@wmcnally thanks for your reply. Am current try some different design for keypoints detection, anchor-free style, but I met some problems:

  1. It looks like Kaopao treat different keypoints as a classfication problem, but I tried directly using a keypoints head to regression the pose, output a. [17, 3] tensor, which can produce keypoint location and scores at once. but now I found the keypoints regression is not very accurate, do u have any idea?
  2. I found my design, whcih using 3 head, keypoints, detect, class heads, make my detection mAP can not goes up. I found it's very intesting since detection should be at least didn't effect so much since I make keypoints head standaredalone. Do u know how to resolve this problem?

my whole arch design based on yolox, and this model acutally worked, but the mAP on detection can't match with KAPAO. I get only about 50 on a yolox-s backbone.