3d_pose_baseline_pytorch
3d_pose_baseline_pytorch copied to clipboard
weigq
coordinate space of GT 2d pose
Hi,
I notice that the 2d pose ground truth in *_2d.pth.tar lies in original image space(1002x1000) after unnormalizing, which is not as expected as cropped image space in 256x256.
Is this the same as the original version?
BTW, I regenerated the hm36 data with cropped 2d pose, and just achieved error of 48mm. I wonder if this difference lead to the performance gap.
THX!
same as the original repo, the 2d pose is obtained from the 1000x1000 image, which means that the unnormalized 2d pose lies in 1000x1000.
Thanks for quick response! So that is to say, the author's implementation in tensorflow does not match what they described in the paper? ( 440x440 -> 256x256 in the paper)
I do not know the details how they generate the 2d detections, however, i think, they crop the image when running the staked hourglass. And the final 2d pose they provide is aligned back to 1000x1000 image.
One more question about the corresponding relation between 2d and 3d joints. Is the 2d input the same as orign repo that includes hip and excludes nose? If so, it's not an one-to-one regression as orign repo. The hip-based 3d pose has an extra nose joint.
in fact, same as the origin repo, the mapping is not one-to-one (the jaw joint only appears in 3D pose), most 3D pose estimation methods also do as this.
@weigq I am wondering that if the inputed 2d pose can be arbitrarily size, which is related with the image size. should I resize the 2d joints into a regular size before I run the inference?
@ky66111 @weigq Could you share the idea of unnormalizing the 2d inputs? I cannot find the mean and std of 2d poses in the given files. Thanks!
@lee-man Were you able to unnormalize the 2d data? I'm having the same issue as you.
@lee-man Were you able to unnormalize the 2d data? I'm having the same issue as you.
@andresherrera97 Sorry that I missed your reply. I can send you the statistics of 2d for unnormalization if you still need it.