LVD
LVD copied to clipboard
Training data + augmentation process for the 3D model LVD_voxels_SMPL
Hi
thank you for your great work!
I wanted to ask what data did you use to train the 3D version of the code LVD_voxels_SMPL with the IF-Net backbone?
Could you also let me know about the augmentation that is done in 3d? The steps from the supplementary indicate that for each scan you:
- select a random walking or running pose from a set of poses (which set?)
- pose the 3D scan by assigning skinning weights of each vertex to those of the closest SMPL point
- manually remove those that get distorted too much
- change the first PCA component to augment the shape (how exactly do you do this for the scan? - you shape the SMPL model and then apply the same linear displacements to the scan vertices by using the closest SMPL point displacements?)
and repeat the process 6 times for each scan.
Did I get it right? Is there something else you do?