inconsistent pose and depth result about sfmlearner-pytorch

[ClementPinard]

Hello,

The gt is only here for Validation, and helps early stopping

If you don't have gt on validation set, I advise you to record a network every 10 epochs or so and then try them all to see which is the best.

The thing is this training workflow is not very robust to overfitting, and quality worsens at some point, the key is to know when it happens.

So instead of testing all network versions on test, I decided to make a vlidation set with gt on which we can choose the best network, so that we don't officially optimize on the test set.