search

facebookresearch / InterHand2.6M

Official PyTorch implementation of "InterHand2.6M: A Dataset and Baseline for 3D Interacting Hand Pose Estimation from a Single RGB Image", ECCV 2020
Other
675 stars 92 forks source link

Why use RootNet rather than gt? #144

Open Divenirelrn opened 10 months ago

Divenirelrn commented 10 months ago

Thanks for your great work! I have one confusion, in this project, label of abs_depth is obtained from rootnet when testing on RHD dataset, and in rootnet project, label of abs_depth is obtained from gt of RHD dataset(z-axis coord). So why not use gt of RHD dataset directly when testing on RHD dataset?

mks0601 commented 10 months ago

Sorry I can't get it. Could you clarify your question?

Divenirelrn commented 10 months ago

Sorry, i didn't clarify my question. I am working on 3D hand pose estimation on RHD dataset, and i don't understand why abs_depth is obtained by RootNet. In this file: https://github.com/facebookresearch/InterHand2.6M/blob/main/data/RHD/dataset.py，between line 78 ~ 84, image , label of abs_depth is obtained from output of RootNet when testing on RHD dataset(line 80), and obtained from GT of RHD when training on RHD dataset(line 84) . Why not obtain label of abs_depth from GT of RHD directly when tesing on RHD, just like training phase?

mks0601 commented 10 months ago

assuming GT information during the testing stage is not fair.

Divenirelrn commented 10 months ago

Ok, I get it. Thank you!