hbb1 / 2d-gaussian-splatting

[SIGGRAPH'24] 2D Gaussian Splatting for Geometrically Accurate Radiance Fields
https://surfsplatting.github.io
Other
1.94k stars 120 forks source link

Mesh extraction within a bounding box #141

Open btoussai opened 1 month ago

btoussai commented 1 month ago

Hello, I need to extract the reconstructed geometry at a high resolution within a bounding box (not the whole scene). How can I do so ? Extracting the mesh at a high resolution from the complete scene is very impractical since it takes a very long time. Thanks.

hbb1 commented 1 month ago

Hi, Note that you can actually get a sdf value at every location (please refer to my implementation of TSDF fusion ). If you want to export a mesh at high resolution, you may compute the sub meshes block by block and finally merge them.

btoussai commented 1 month ago

Thanks, I've managed to add bounding box corners to the mesh extraction, both in the bounded and unbounded case. I'm observing several copy of my object due to some points of views having incorrect average depths estimates. Using the median depth for mesh extraction solves the issue but introduces noise. Is there any advantage or disadvantage to training with the median depth instead of the average depth ?

hbb1 commented 1 month ago

Median depth is more robust to outliers but has the issue of "disk-aliasing". You may see holes in the mesh when we don't have sufficient primitives to represent the shape.

btoussai commented 1 month ago

I re-trained on my dataset with the median depth and it solves some problems appearance-wise but it creates holes in some cases, as you mentioned. In any case, the mesh extraction is still better than with the average depth, although it is more noisy.