statistical outlier removal isn't enough

[soumya997]

Hi guys, I noticed there's a new denoising feature available.

Previously, I used statistical outlier removal from Open3D on the extracted ply file, which wasn't much effective, struggled with dense floaters generated in 3DGS. These floaters are usually very dense, which I believe makes it difficult for general algorithms to remove them effectively.

I was wondering if there might be a better approach to tackle these floaters. If they can be successfully removed, converting the point cloud file back to a 3DGS file should also eliminate the floaters there. Below are some examples where I'm encountering this issue.

img 1	img 2

[Lewis-Stuart-11]

Hi, I agree that the statistical outlier removal is not effective enough, mainly due to the fact that once all the points are sampled for large noisy Gaussians, it is hard to identify them for removal.

I have now implemented a new visibility threshold, which now culls Gaussians that do not contribute the required amount to the rendered images of the scene. This significantly reduces the amount of noise in the scene. Let me know if this works better for you