About best result

[google12306]

First of all, thank you very much for the excellent work！
Secondly, I would like to know how to determine which optimal result should be loaded for next stage, and what are the metric for choosing, D1, EPE or another?
Lastly, the paper mentioned implementation details regarding training on the KITTI dataset for 500 times separately , is the learning rate still 0.001 at that time? Is the validation set still 20 images? How are the optimal weights to be loaded determined on the mixed KITTI 2012 and KITTI 2015?

I look forward to your response.Thanks!