search

aeolusguan / NMRF

[CVPR 2024] Neural Markov Random Field for Stereo Matching
MIT License
101 stars 5 forks source link

Possibility Adapting NMRF into a Self-Supervised Network. #5

Open Magicboomliu opened 5 days ago

Magicboomliu commented 5 days ago

Hi! Outstanding works.

I am writing this issue just to discuss whether we can use the NMRF in a self-supervised manner without using the ground truth disparity.

The main question is how to apply the Disparity Proposal Loss in a self-supervised manner since it requires the GT disparity for matching. Thanks a lot if you are interested.

aeolusguan commented 3 days ago

Thanks for your interest in our work. If Disparity Proposal Loss is your main concern, the Disparity Proposal Network could be replaced with another NMRF inference stage, which induces only slight performance drop in the supervised setting. Good Luck! :muscle: