Severe Surface Spikes and Blurring in 3DGS Reconstruction

[SimonWang404]

Hi, thank you very much for your excellent work!

I am currently working with a 3D Gaussian Splatting (3DGS) pipeline for scene reconstruction, and I have encountered several issues related to surface quality after reconstruction.

Observed Issues

• The reconstructed surfaces exhibit spike-like or needle-like structures.
• There are noticeable “smearing” or blurry effects on surfaces.
• Geometric edges lack sharpness and appear overly smooth or distorted.

Questions

1. What are the main causes of these artifacts?

• Insufficient training or lack of convergence?
• Distortion from fisheye images?
• Improper learning of Gaussian scale parameters?

2. Are there recommended solutions at the training level?

• Applying regularization on Gaussian scales?
• Preventing extreme anisotropic Gaussians during optimization?

3. Regarding fisheye inputs:

• Is it necessary to undistort images before training?
• Are there recommended pipelines specifically designed for fisheye data?

4. Can these issues be effectively fixed via post-processing on the exported PLY (e.g., filtering or adjusting Gaussian parameters)?
   Or is retraining required to fundamentally resolve them?

Any suggestions or insights would be greatly appreciated. Thanks again for your work!

[SimonWang404]