search

xingyizhou / pytorch-pose-hg-3d

PyTorch implementation for 3D human pose estimation
GNU General Public License v3.0
615 stars 141 forks source link

the 3d loss #69

Open JWSunny opened 5 years ago

JWSunny commented 5 years ago

hello：

Thanks for your sharing about the work. In the code(such as h36m.py and train3d.py), i don't understand the 'reg_ind' and 'reg_target', can you explain the parameters and why use them to calculate the 3d pose loss, is there any bais or reference? thanks.

xingyizhou commented 5 years ago

Hi, Thanks for looking at the code. These are good questions.

  1. output[-1]['depth'] is in shape batch x 16 x H x W. The code first extract the value at the ground truth locations (batch['reg_ind']) to convert it into batch x 16 x 1. And then they are the same as directly using regression. The code is from CornerNet.
  2. varloss is the eq. 3 in the paper.
JWSunny commented 5 years ago

@xingyizhou In other words，the locations (batch['reg_ind']) values and the regression depth(batch['reg_target']) have some relationships? from the paper: ConerNet?

xingyizhou commented 5 years ago

Yes.

JWSunny commented 5 years ago

@xingyizhou hello, in the code h36m.py, self.mean_bone_length=4000, but i saw other places are possible 4200mm, i want to how to calculate the value; for example, (x,y,x), x, y are the locations of pixels in the RGB image or in the real scene, but the z is the depth value of the real scene; can you give some explanations, thanks.