search

kcheng1021 / GaussianPro

[ICML2024] Official code for GaussianPro: 3D Gaussian Splatting with Progressive Propagation
https://kcheng1021.github.io/gaussianpro.github.io/
MIT License
551 stars 36 forks source link

Qustion about the finall point cloud #44

Open ZhuoxiaoLi opened 1 month ago

ZhuoxiaoLi commented 1 month ago

Hello, Thank you for your excellent work!! I tried using the GUGONG demo you provided for code testing, and I found that after 30,000 rounds of training, a small part of the point cloud was greatly enhanced (see the figure I provide below), but only a small part. The remaining part was compared with the original 3DGS after the point cloud was densified, and it was found that the point cloud densification had little difference. Still, the PSNR and other test indicators were improved. I would like to ask: 1. Is this because the current version of the code is only the initial version?

  1. Is there any other function that has not been put into the code? gugong
kcheng1021 commented 1 month ago

Hi, thanks for your great question! The improvement could come from both densification and planar constraint. I guess the reason that the specific part has much denser point cloud is that 3DGS predicts wrong depth in that region.

For your question, there are no core function that missed in the current version. The next version is mainly for the acceleration and code cleaning.

ZhuoxiaoLi commented 1 month ago

Thank you for your prompt reply!

I checked several input images and the point cloud at the end of the training. Find that this point cloud part may not be caused by error. Maybe it's a very complete densified point cloud. It completely restores the entire roof structure, with more integrity than the other parts.

I would like to ask since this small part can achieve such a high point cloud density, is there any way to make other parts can also achieve such a high point cloud density?

image image

kcheng1021 commented 1 month ago

Thanks for your insightful question and observation! The densification of the point cloud is affected by both the propagation process and the 3DGS's original rendered depth. If 3DGS fails to densify a region, then the propagation will compenstate the region if it estimates the geometriclly consistent planes. That is indeed what can be further explored for 3D reconstruction.