search

autonomousvision / sdfstudio

A Unified Framework for Surface Reconstruction
Apache License 2.0
1.96k stars 183 forks source link

Has anyone managed to reproduce SOTA mesh results with neus-facto? #253

Open juuso-oskari opened 10 months ago

juuso-oskari commented 10 months ago

Hi,

I would like to know if anyone has managed to get produce SOTA chamfer distance results with neus-facto? I try to train the model with the command: 

ns-train neus-facto --pipeline.model.sdf-field.inside-outside False sdfstudio-data --data data/dtu/scan65

And I let it run 30 minutes (on Nvidia A4500 GPU). Then I use the monosdf evaluation procedure to get a culled mesh and evaluate it against the groundtruth points. But the culled mesh I get looks like this: image

PSNR and the chamfer distance are way off the SOTA: 19.4 & 2.14. For example monosdf reports to get chamfer of 0.78. Is it simply so that using these fast samplers like proposal-sampler the mesh is just not that good?

I would desperately need some model to produce SOTA mesh results so I could test my module on it.

niujinshuchong commented 10 months ago

Hi, your result look quite different from the one we got with neus-facto, please check the video of the rendered mesh in readme for comparison. I am not sure what cause the differences.