rchi-lab / bodymap Goto Github PK

Hello Abhishek! Congradulations for your new CVPR! I am very interested in BodyMAP but I am still confused by the pipeline of BodyMap.

In the paper, you mentioned the importance of getting the appiled 3D pressure map and proposed a network to predict 3d human mesh and 3D pressure map simultaneously. However, as the 3D pressure map ground truth is generated (projected) from the 2D pressure map collected by a pressure sensing mat (or simulation) and the 2D pressure map is one of the network inputs, why do we need to predict the 3D pressure map instead of simply predicting the human mesh and then projecting the 2D pressure map input to the human mesh in the same manner to obtain the 3D ground truth (which is a similar pipeline as Bodies at Rest and Bodypressure but has both input modalities)? If the pressure distribution of predicted 3D pressure map is not consistent with the 2D input ( which is prone to occur), which one should I believe in?

In short, given that the 3D pressure map ground truths are generated from 2D pressure map during model training, why don't we directly generate 3D pressure map outputs from 2D inputs instead of predicting them through the network?

Thanks very much for your reply!