How to define 'visible' in current training views

[foricee]

Thank you for your great works! While I have something not clear. Supplemental Material, at the end of the SDF Regularizer part, the paper writes:

Specifically, for a triangular face f extracted from tetrahedron T, if f is not visible in current training views, we encourage the SDFs` at vertices of T to be positive with BCE loss

My question is that how to define 'visible' in current training views. I've read the code, but still can't find the implementation. buffers['occlusion'] in dmtet.py seems like to be relative. But I'm not sure. Could you give me some other supplemental material or advices?

[foricee]

After reading https://nvlabs.github.io/nvdiffrast/, I think I've found the solution

These codes are added in render/render.py, at the end of function render_mesh(), https://github.com/NVlabs/nvdiffrec/blob/main/render/render.py#L251

    rast = layers[0][1] 
    tri_id_col = rast[..., 3]
    visible_tri_id = torch.unique(tri_id_col[tri_id_col > 0]).long() - 1
    visible_tri = mesh.t_pos_idx[visible_tri_id]

Is this the right way to extract all visible triangles?

[frankshen07]

Thank you for your interest in our work! Yes, the way you extract visible triangles is correct. Another thing you should need is https://github.com/NVIDIAGameWorks/kaolin/blob/6569ea3e89f3f70ec7885ec053d900653ade67f3/kaolin/ops/conversions/tetmesh.py#L105 which maps visible faces to tetrahedra.

[foricee]

Thank you for your interest in our work! Yes, the way you extract visible triangles is correct. Another thing you should need is https://github.com/NVIDIAGameWorks/kaolin/blob/6569ea3e89f3f70ec7885ec053d900653ade67f3/kaolin/ops/conversions/tetmesh.py#L105 which maps visible faces to tetrahedra.

It works. Thank you!