search

Jeff-sjtu / HybrIK

Official code of "HybrIK: A Hybrid Analytical-Neural Inverse Kinematics Solution for 3D Human Pose and Shape Estimation", CVPR 2021
MIT License
1.17k stars 142 forks source link

Traning with RLE loss #91

Closed PGogo closed 1 year ago

PGogo commented 2 years ago

Thanks for your great work and sharing your implementation.^^ I noticed that another work "Human Pose Regression with Residual Log-likelihood Estimation" is also one of yours. And it is superior to the heatmap-based methods in a regression manner. I wonder if it still works in HybrIK. And what were the results between heatmap-based methods and regression-based methods in the experiment, of course if you ever gave it a try.

Jeff-sjtu commented 1 year ago

Hi @PGogo, thanks for your interest in our work. Actually, we have tried RLE on HybrIK and found the performances (MPJPE) on 3DPW and Human3.6M are similar to the heatmap-based version. But RLE has the potential to defeat the heatmap-based version in the sciences of truncation and occlusion.

PGogo commented 1 year ago

Thanks!