densebody_pytorch
densebody_pytorch copied to clipboard
Lotayou
some tips(advice) to better performance and UV-map performance
Hello, i have tried your code and make some changes(tips) to perform better. Hope it would help：
- L1 loss weight . The L1 loss weight generated is wrong accrodding to your code and your uv-map-template after dilation. UV map shouldnot be dilated
- TV loss. I added tv loss with corrected weights. The TV loss weights should smaller than L1 loss two order of magnitudes . The results showed it improved a little.
- From UV map to Verts. changed SAMPLE algorithm little.
- Increase joints-center verts loss weight.
However，the MPJPE-PA i calculated is about 62, much lager than the paper listed. The MPJPE-PA between joints from UV-map-generated-form-gt3d and gt3ds is about 22mm, it also much larger than paper listed.
@GuodongQi would you like to share your modified code? And how to calculate the 3D joint? SMPL fitting or just multiple the joint regressor?
@willie1997 I am afraid not. The latter，just multiple the joint regressor.
@GuodongQi Thanks for the tips. Do you care to make a pull request? Also I think it's perfectly normal for my implemetation to be worse than the paper reported, since I only used 1/100 of the h36m dataset (i.e. about 30K frames). If you have access to the full human36m data you can train a much better model.
