search

anibali / margipose

3D monocular human pose estimation
Apache License 2.0
98 stars 20 forks source link

zy and xz axis permutations swapped #22

Closed tnybny closed 4 years ago

tnybny commented 4 years ago

If I'm reading this right, the order of permutations is swapped between the zy permutations and xz permutations. That is, the code snippet claiming to perform axis permutation for zy axes is actually doing the corresponding permutation for xz axes, and vice versa. Since the targets target_zy and target_xz are being extracted in the correct order, this could have an impact while calculating the pixelwise heatmap losses. https://github.com/anibali/margipose/blob/e96d59187dc17651ab184ca263f9a1a150cfa201/src/margipose/models/margipose_model.py#L94-L97

anibali commented 4 years ago

I believe that the code here is correct.

mid_in has axes in BCHW order. After permuting, mid_out (zy) will be BWHC (henceforth treated as BCHD) and mid_out (xz) will be BHCW (henceforth treated as BCDW).