Mesh Holes Issue

[muzol132126]

I followed the README for GS training and mesh extraction on DTU scan24. With antialiasing on in GaussianRasterizationSettings, the output mesh has many holes. I checked steps but no luck. Any ideas on what could be wrong?

[muzol132126]

Besides, some face normals of the mesh are flipped, and there are convex hulls forming which make originally open surfaces closed. I suspect it's an issue with extracting mesh from SDF using Marching Cube as I encountered similar problems in the Neural-Pull method too. Any insights or solutions would be much appreciated.

DTU scan40 GS-Pull result and Ground Truth

[wen-yuan-zhang]

We shrink the double-layer SDF surfaces inward to the zero iso-surface, so there exists two layers of opposite surfaces on the zero level-set, which is a normal phenomenon. If you enable the "double back-face" option in meshlab, you can see consistent oriented faces.
I'm not quite sure why this issue occurred. Did you use our released dataset and follow all the training steps provided in readme.md?

[muzol132126]

Thank you for your reply. It has made me aware of the normal flipping issue.
I followed the training steps in the readme. Moreover, I have tried the TSDF fusion method to extract the mesh, but there are still holes.

[muzol132126]

Thank you for your reply. It has made me aware of the normal flipping issue.
I followed the training steps in the readme. Moreover, I have tried the TSDF fusion method to extract the mesh, but there are still holes.