search

elliottwu / unsup3d

(CVPR'20 Oral) Unsupervised Learning of Probably Symmetric Deformable 3D Objects from Images in the Wild
MIT License
1.19k stars 192 forks source link

How to customize a higher resolution model? #9

Closed stone100010 closed 3 years ago

stone100010 commented 4 years ago

How to customize a higher resolution model?

elliottwu commented 4 years ago

Hi!

You could try to increase the output resolution of the networks by adding more upsampling layers. A couple of potential issues might arise when you try to increase the resolution to, say, 256x256:

  1. Higher resolution depth maps may result in crazy occlusion and the gradients through the renderer will be too noisy to learn anything meaningful, especially when the depth maps are noisy initially from randomly initialized network weights. But it is possible to add a smoothness loss on the depth maps.
  2. Training will much slower, mainly due to the rendering cost.
  3. The simple Lambertian model may not be powerful enough to render photorealistic images, even if it has a higher resolution.

All of these are good research problems that can be tackled in future work.