search

ali-vilab / Infusion

Official implementation for paper: InFusion: Inpainting 3D Gaussians via Learning Depth Completion from Diffusion Prior
MIT License
534 stars 16 forks source link

Why train incomplete 3D Gaussians first？ #12

Open Dumeowmeow opened 5 months ago

Dumeowmeow commented 5 months ago

What is the purpose of doing this？

Johanan528 commented 5 months ago

Our goal is to construct an incomplete Gaussian. According to our experiments, some segmentation methods at that time could not segment Gaussian robustly, so we directly trained an incomplete Gaussian. Of course, this is not the core part of this project. Now, perhaps some Gaussian segmentation methods have better performance.

Dumeowmeow commented 5 months ago

Thank you for your reply.But after I read the train.py in gaussian-splatting, I found that if mask_training is True, the loss is computed between masked render image and masked gt-image,I can't understand why does this allow us to train incomplete Gaussians.And another question is, does an incomplete Gaussian mean a Gaussian with the inpaint part gaussians completely removed?