hbb1 / 2d-gaussian-splatting

[SIGGRAPH'24] 2D Gaussian Splatting for Geometrically Accurate Radiance Fields
https://surfsplatting.github.io
Other
1.94k stars 120 forks source link

gaussians spread out at no background dataset #22

Closed qianqiuzzz closed 2 months ago

qianqiuzzz commented 4 months ago

When I use a dataset with no background, gaussians begins to become scattered, they appear in the background position after 7k iterations, causing black areas to appear in the mesh. Is this phenomenon caused by Normal Consistency regularization? Is there any way to avoid this problem? image image

hbb1 commented 4 months ago

Hi, We haven't tested our algorithm for scenes without background. Because the floater can learn fake background color, an intrinsic ambiguity of volume rendering, the normal consistency loss may treat it as solid object and may spread these floaters. I suggest you move the start iteration of normal consistency after densitification , e.g., 15k, when the floaters may have been effectively cleaned by reset opacity. Or if you have a background mask, just use it to mask the regularization loss. Hope these will help.

shiwangyan commented 3 months ago

Hi, We haven't tested our algorithm for scenes without background. Because the floater can learn fake background color, an intrinsic ambiguity of volume rendering, the normal consistency loss may treat it as solid object and may spread these floaters. I suggest you move the start iteration of normal consistency after densitification , e.g., 15k, when the floaters may have been effectively cleaned by reset opacity. Or if you have a background mask, just use it to mask the regularization loss. Hope these will help.

怎么移动normal consistency的迭代呢?