Question about the silhouette

[l1346792580123]

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]

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