search

garyzhao / SemGCN

The Pytorch implementation for "Semantic Graph Convolutional Networks for 3D Human Pose Regression" (CVPR 2019).
https://arxiv.org/abs/1904.03345
Apache License 2.0
465 stars 78 forks source link

Do you ever have tried cropping the joint? #11

Open gh18l opened 4 years ago

gh18l commented 4 years ago

hello! Do you ever have tried cropping the image into patches with person-centered to reduce the influence of bias of 2D joint? Does it probably have any improvement? Thank you!

garyzhao commented 4 years ago

Hi @li19960612 ,

Thanks for your interest in our work.

It's a good question. If we crop the image into a patch and resize it, we will lose the camera information. In this case, if we want to evaluate the performance with MPJPE, we have to store the camera information and then use it during testing to reconstruct the output with the correct scale. This means we will use the ground truth information during testing, which is not a fair comparison to other methods.

See one related case here: https://github.com/JimmySuen/integral-human-pose/issues/30

Best, Long

gh18l commented 4 years ago

@garyzhao ok, I see. Thank you for your reply!