search

YoYo000 / MVSNet

MVSNet (ECCV2018) & R-MVSNet (CVPR2019)
MIT License
1.38k stars 303 forks source link

How to deal with the surface that has high reflection? #94

Open fred2019-dev opened 4 years ago

fred2019-dev commented 4 years ago

image image image

for example, for the surface like the bottom part in the above images, the reconstruction result using currently provided pretrained model is not very acceptible.

Could such cases be addressed by training the network using similar data? Has anyone tried yet?

tatsy commented 4 years ago

I think handling shiny materials is a common problem of the multiview stereo, while current deep learning techniques alleviated the problem.

The DTU dataset includes some shiny materials (such as No. 3, 75, 99, and 110), so I think the depth estimation for such shiny materials could be learned a bit. However, the results are highly dependent on the training data and I do not think all the scenes in the world can be handled properly.

Hope it helps.