About data-processing in PoseBert

[buptxyb666]

Thanks for your great works in HMR area.

For data processing in PoseBert, the paper mentions that some canonicalizations are made on training data. I've noticed that these canonicalizations cannot be applied during inference to ensure image-mesh alignment. Would this difference in data distribution affect performance? Looking forward your reply. Thank you so much.

[fabienbaradel]

Hi @buptxyb666 , thanks for your interest in our work.
In PoseBERT we are taking as input a sequence of estimated SMPL poses (or 3d poses) of a single person as input and we want to refine the estimated SMPL poses.
However in this work we are taking as input a single RGB image and we want to detect persons and estimate their properties (distance from the camera, SMPL-X shape, SMPL-X pose). Since we are directly regressing the SMPL-X parameters for each detected person (without taking 2d pose into account) we cannot ensure an image-mesh alignment.
I hope that my answer was clear enough,