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
460 stars 78 forks source link

About the global offset #12

Closed sjtuxcx closed 4 years ago

sjtuxcx commented 4 years ago

Hello, I noticed you remove the global offset in data_utils.py but it will lose the depth information. It will make a difference with others in the evaluating stage because you set the hip joint as the original point but not the camera. Your estimation will be easier because you don't need to estimate the depth. I do not understand. Can you help me?

garyzhao commented 4 years ago

Hi @sjtuxcx

Thanks for your interest in our work.

We follow the setting of [1] to do the prediction, which indicates that "Since we do not predict the global position of the 3d prediction, we zero-centre the 3d poses around the hip joint (in line with previous work and the standard protocol of Human3.6M)." (The first paragraph of Section 3.2 in [1])

In addition, if we evaluate the prediction by MPJPE (Protocol 1), the first step is to align the root (central hip) joint of the prediction and ground truth. I do not think predicting the global offset makes any difference here.

[1] Martinez et al. A simple yet effective baseline for 3d human pose estimation. ICCV 2017.

Best, Long