River-Zhang/SIFU

Table 6 in Table

ZZWENG opened this issue · 3 comments

ZZWENG commented

Hi authors, could you provide some details on your evaluation protocol for Table 6? Specifically,

  1. what is the rendering size?
  2. how many views per subject?
  3. which backbone for lpips?

Thank you very much!

River-Zhang commented

Thanks for reading our paper.

  1. The rendering size is 512 as the size of GT images in the dataset is 512.
  2. We use four views per subject, by rotating {0, 90, 180, 270}. This is the same as our normal loss calculation.
  3. We use VGG backbone for lpips calculation.
ZZWENG commented
River-Zhang commented

Hi, thank you for considering our work for your ECCV submission! Unfortunately, the texture inference code and the original qualitative results from Figure 5 are not currently available for public release. However, the coarse texture prediction is contained in our released code. You can try that by uncommenting the code in apps/infer.py line 709 and 717, use gen_mesh_color() to get the coarse texture.

         final_colors=gen_mesh_color(verts_pr, model.netG, device, in_tensor)
         final.visual.vertex_colors = final_colors

Also, you can use the quantitative results in the paper as a reference. We haven't released the texture refinement stage because it needs GPT-4V and it is not available for everyone. We are looking for alternative ways for a stable, cheap, and effective refinement.
Best wishes for your submission!