Is it possible to train the DispNet alone?

[sanweiliti]

Hi,
Since the depth net and pose net are trained jointly, they must benefit from each other. Is it possible to only train the DispNet alone as a supervised problem without using pose information? I understand that one of the highlights of your paper is to use unsupervised learning, but I'm curious is it reasonable to train the DispNet alone in a supervised framework? I'm studying different depth estimation network structures, and it will be more fair to compare all structures in a similiar supervised framework. Thank you!

[bachw433]

hi sanweiliti,
you can check this paper:
https://arxiv.org/abs/1512.02134
dispNET comes from there and it was a supervised model.
(actually, all the model can be either supervised or unsupervised only depends on your loss fucntion)

and BTW, I'm interesting in your study. Which network is the best one so far XD?

[sanweiliti]

Hi @bachw433
Thanks! I will try to modify it. My study focuses more on relationships between depth estimation networks and segmentation networks so I can't really say which model is the best.