search

hbb1 / 2d-gaussian-splatting

[SIGGRAPH'24] 2D Gaussian Splatting for Geometrically Accurate Radiance Fields
https://surfsplatting.github.io
Other
2.14k stars 149 forks source link

Question about synthetic dataset #50

Closed Yubel426 closed 6 months ago

Yubel426 commented 6 months ago

Hi! When trying to reconstruct the mesh on the synthetic, I noticed a significant presence of black erroneous faces in the output. Have you tested on this dataset before? Should I consider additional hyperparameter adjustments? Any insights or suggestions you could offer for this issue would be greatly appreciated. 1716974362113 1716974419046

hbb1 commented 6 months ago

I need further clarification. Is there floaters outside the objects, i.e., 2D Gaussians that fake the background color?

Yubel426 commented 6 months ago

No additional Gaussian points are generated outside the object (almost none), which I find weird. Below is the corresponding point cloud. 1717035348526

hbb1 commented 6 months ago

I can see some outliers in free space, I am not sure if this causes the issue. Maybe similar to #22 , it can produce gaussians with fake background color when introducing depth-normal consistency.

Yubel426 commented 6 months ago

Thanks, it does seem to be caused by normal consistency.

543877815 commented 4 months ago

How to solve this problem?

Yubel426 commented 3 months ago

@543877815 You can remove the normal consistency loss or introduce the mask regularization. Reference to https://github.com/hbb1/2d-gaussian-splatting/issues/22 and https://github.com/hbb1/2d-gaussian-splatting/issues/120