search

nv-tlabs / LION

Latent Point Diffusion Models for 3D Shape Generation
Other
735 stars 57 forks source link

Can 3D objects with holes or imperfections be used for training the model? #71

Open Javiergonb opened 2 months ago

Javiergonb commented 2 months ago

Hi @ZENGXH ,

I have been trying (and failing), to train this model using a dataset of 3D objects that have holes or imperfections, such as museum objects of ancient pottery. The results are not very impressive and I am starting to suspect it might have something to do with that. So i have the following questions:

1.- Will using 3D objects with holes or imperfections negatively impact the training process?

2.- Are there any recommended preprocessing techniques to handle these imperfections before using the objects for training?

I appreciate any guidance or insights you can provide on this matter. Thanks!

ZENGXH commented 2 months ago
  1. yes they will have negative impact. The generated shape will also have holes and the imperfections.
  2. some data cleaning may help, for example, run a surface reconstruction method first to remove those hole and imperfections, then sample points on the reconstructed surface.