Remove floaters

[pierotofy]

There should be a way to reduce/filter floaters.

These significantly reduce the quality of visualization.

I thought about doing a nearest neighbor search from of the splat means compared to the points in the initial point cloud and filter any splat that is not within X units of an original point. But I don't know if this would work, as certain areas could have no initialization points. If there are better/proven methods I'd like to hear it.

[pfxuan]

This method can give some help. But it’s a little computation heavy.

[pierotofy]

Interesting, it's similar to this one from PDAL: https://pdal.io/en/2.7.0/stages/filters.outlier.html#statistical-method (which can be made to run fast).

[pfxuan]

Good idea. Open3D also has a few similar built-in functions: https://www.open3d.org/docs/latest/tutorial/Advanced/pointcloud_outlier_removal.html

[pierotofy]

I've tried to implement this method in https://github.com/pierotofy/OpenSplat/tree/filter, but although there's some reduction of noise, there's also an increase in other parts (at least in the train dataset) near the object of interest. The validation loss also seems to increase.

[pfxuan]

Perhaps removing floaters could be considered as a post-processing step after the 3DGS model gets fully trained. Additionally, I found enhancing the training code, improving camera pose, and removing blurry image could substantially minimize floating artifacts.