autonomousvision / gaussian-opacity-fields

Gaussian Opacity Fields: Efficient and Compact Surface Reconstruction in Unbounded Scenes
https://niujinshuchong.github.io/gaussian-opacity-fields/
Other
559 stars 26 forks source link

Can I extract mesh from trained vanilla 3DGS model? #61

Closed Runyu-Zhou05 closed 3 days ago

Runyu-Zhou05 commented 1 week ago

Hi, thanks for your great work! I wonder if it's possible to extract mesh from a trained vanilla 3DGS model so that I don't have to train it again using GOF. When I run mesh extraction directly it says filter_3D is missing in the point_cloud.ply file. Please let me know if I just need to do some minor modifications to the code. Thank you!

niujinshuchong commented 4 days ago

Hi, Yes as long as the underlying representation is 3D Gaussians. You could set the filter_3D to zero for that.

But the original 3DGS codebase suffers from some artifacts (see our mip-splatting project https://niujinshuchong.github.io/mip-splatting/) and we fixed it in mip-splatting. This project GOF is built on mip-splatting's solution and there is some inconsistencies to the original 3DGS's rendering, therefore the results might not be very good. I would suggest you train your model with mip-splatting and extract the mesh with GOF or retrain your model with GOF.