Approximate to my own garment?

[chenbomath]

Hi, thanks for your sharing. Can I use it to approximate new garment mesh instead of that recovered from the latent space? If yes, what is the requirement of the garment meshes? And where should I change the code to make it work?

[bguillard]

Hello, The pretrained network that we share was only trained on 4 different items of garment, and thus has limited generalization capabilities.

If you want to generalize to new garments, you need to train on a larger dataset that covers more items. For this, you must:

1. Normalize your meshes (to a unit sphere for example), and generate UDF samples for them.

2. Train a point cloud encoder + UDF decoder, as we do in DrapeNet, or an auto-decoder as in DeepSDF (but with UDFs this time).

I hope this helps, Benoit