Marching Cubes Quality

[cheind]

Hey,

thank you very much for your wonderful contribution!

Issue

We started testing NeuS2 and found that the mesh quality generated from within the UI seems inferior to the results we get from loading the density image and processing it in external marching cubes or dual contouring algorithms. Which is suprising to me, since the density images loses information in part to the mapping from floating point SDF to discrete intensity values. Note, this issue might exist in instant-ngp code basis as well, but we haven't tested.

Results

The following images are generated as follows:

• load the lego scene in NeuS2
• let it run till convergence
• stop the training
• set the sdf grid res to 64, range to 1
• export the mesh MESH_A and density image
• process density image with sdftoolbox to generate mesh MESH_B (marching cubes) and MESH_C (dual contouring / surface nets)
• load the meshes in Meshlab and take screenshots from the same perspective

MESH_A:

MESH_B:

MESH_C:

Any feedback is welcome :)

[cheind]

might be related to #17

[19reborn]

Hi @cheind, this is very interesting. Have you tried higher resolution (e.g. 400)?

[cheind]

Hey @19reborn, no we haven't done that yet (at least not extensively). My subjective feeling is, that with higher resolution the error becomes less apparent.

We plan to run reconstruction tests on real/synthetic scenes for which a ground-truth surface model exists and then measure the reconstruction error by comparing surface meshes directly (i.e Hausdorff distance).