Profactor / continuous-remeshing

Continuous Remeshing For Inverse Rendering
MIT License
445 stars 24 forks source link

Question about the silhouette #7

Open l1346792580123 opened 1 month ago

l1346792580123 commented 1 month ago

Thanks for sharing this excellent work! I have found that in example.py only normal loss is used to optimize the mesh. I am curious why using only normal loss without silhouette loss allows the sphere to expand to the target mesh. And I have noticed that the paper mentioned that there are limitations with silhouette gradients. Can you explain it in detail? Does adding silhouette rendering loss make the results worse?

brotherofken commented 2 days ago

Renders and targets have four channels. The last channel is a mask, which is used in L1 loss.