search

3dpose / GnTCN

Graph and Temporal Convolutional Networks for 3D Multi-person Pose Estimation in Monocular Videos
MIT License
96 stars 13 forks source link

2 different checkpoints for GCN #12

Open edricwu opened 3 years ago

edricwu commented 3 years ago

Thank you for sharing your work. I have one question:

  1. There are 2 checkpoints for GCN, model_gcn and model_gcnwild. What is the difference between the two? I notice that model_gcn is used for H36M evaluation and model_gcnwild is used for MuPoTs. Is it because H36M and MuPoTs have different keypoints definition?
3dpose commented 3 years ago

Thanks for your interest in our work.

The major reason there are two checkpoints for GCN (model_gcn and model_gcnwild) is that model_gcn is trained on H36M training set only, and model_gcnwild is trained on training+testing sets of H36M. A second reason is that model_gcn used unnormalized skeleton and model_gcnwild used normalized skeleton because the normalized skeleton setting is more suitable for wild videos and the unnormalized skeleton works better on H36M dataset.

edricwu commented 3 years ago

Thanks for the prompt reply. By normalization, do you mean scaling the average joint-to-pelvis distance to 1?