r0nn13/learn2rec

learn2rec's Issues

Any update?

When will it be released.

Very nice paper! Looking forward to the code.
When can we try it?

"Check back soon for more details." It has been almost 1 year...

A discussion of paper

Nice project!
I have some questions on the paper and experiments. Since the code is not available yet, I post the questions here and hope to get some discussions.
First is the photo-consistency assumption used to calculate the gradient and Jacobian. Since many images do not hold the assumption:

How does this affect the network?

Second, although the learned optimizer can regularize the depth map, how the depth of unobserved areas be estimated when there are no two-view observations?
As shown below, the net can deal with the situation, but why it can?

I guess the learned structure prior plays an important role here.

Third, can you explain (or give out the code) Section 7.1 of the paper? As I am understanding, the trained net can fit all four functions (Eq.14～Eq.17). During the test, one of the functions is sampled and the net can fit it according to the data. If so, how do you generate the gradient? Since you do not know which function is sampled.

Lastly, thank your nice work and I hope to see the code soon!