per shape or per category normalization ?

[lx709]

Hello @yccyenchicheng , Thanks for sharing the code of your wonderful work. In the original DISN code, there are two normalization ways, per-shape used in create_point_sdf_grid and per-category used in create_point_sdf_fullgrid.py. From your code, it seems you are using per-shape normalization before calculating SDF values, right? Why did you choose the per-shape one? Could you please elaborate more on your choice?

[yccyenchicheng]

We want to normalize such that the shape is located at the center (0, 0, 0) to ease the training of AE and the transformers. I think a analogy in 2D face image synthesis is the alignment with the eyes. It would be great if we can validate if this really helps the training.

I hope that makes sense!