Closed sun510001 closed 1 year ago
Thanks @sun510001 !
Increasing the Laplacian regularizer would have been my first tests as well. You can test the
"lock_pos" : true
option in the config, which disables geometry optimization in the second pass .
You can also test decreasing the learning rate in the second pass, like
"learning_rate": [0.03, 0.003],
or similar.
Also, nvdiffrec(mc) is quite sensitive to inaccuracies in the foreground segmentation masks and poses. Your reconstructed silhouettes look a bit uneven, so if there is a way to provide better masks with your dataset, that could likely help quality a bit here.
@jmunkberg Thanks for your reply.
I checked the previously generated meshes and datasets. The datasets with good matting generated fewer fragments in the textures of the mesh.
Hello, Firstly, thanks for your excellent work. I'm trying nvdiffrec and nvdiffrecmc on my own data.
It works pretty good at first pass:
![image](https://user-images.githubusercontent.com/32349603/208570390-304e504c-59d9-4725-8d43-69100d9bba19.png)
but lots of wrong fragments are appeared after second pass:
![image](https://user-images.githubusercontent.com/32349603/208570335-0279a429-b1f9-4ed5-a96c-acc1b08637e0.png)
This problem occurs with both nvdiffrec and nvdiffrecmc. Lock pos can solve this problem, but I want to use regularizer to make mesh smoother. I have tried laplace_scale 3000 6000 and 10000; laplace: absolute and relative, but wrong fragments still appeared after second pass.
Any input here would be greatly appreciated, thank you!