search

autonomousvision / gaussian-opacity-fields

[SIGGRAPH Asia'24 & TOG] Gaussian Opacity Fields: Efficient Adaptive Surface Reconstruction in Unbounded Scenes
https://niujinshuchong.github.io/gaussian-opacity-fields/
Other
683 stars 34 forks source link

Bad mesh with marching tetrahedra on DTU #19

Closed brianneoberson closed 4 months ago

brianneoberson commented 5 months ago

Hello,

First of all, thank you for your work! I wanted to test your code on my own dataset and started by testing it on the dtu scene 37 and I got very bad results using the marching tetrahedra mesh extraction. I ran the run_dtu.py script only on this scene and I got the following results: This is the mesh result with TSDF: dtu37_gof_snapshot01

And here with Marching tetrahedra: dtu37_gof_binary_search_snapshot02 The scene is not recognizable at all... Do you know why this is the case? I was under the impression that marching tetrahedra should give better results than tsdf.

Also, it took over an hour to train on a NVIDIA RTX A6000 where as in the paper you mention it is about 30 minutes. Did you use specific training parameters for the results you got in the paper?

Best regards, Brianne

Xxxy13 commented 5 months ago

Hi, Brianne. I have encountered the same problem as you. Have you found the reason?

niujinshuchong commented 5 months ago

Hi, you need to choose double face inside meshlab.

image

The extracted mesh contains large triangles from the background and our simple filtering strategy based on edge distance could not filter this in some cases. If you only care about the foreground region, you can define a bounding box and only select the gaussian inside the predefined box to create the mesh.