Stereo + Tempora (training with stereo video like monodepth2)

[Ecalpal]

Hi, @mdfirman @JamieWatson683

I see that you claim to have conducted relevant experiments in issue#12.

I modified the mono_dataset.py and trainer.py according to Monodepth2, trained the manydepth from scratch (with pre-trained resnet), and got an unsatisfactory result.

-> Loading weights from ./log/manydepth_ms/models/weights_19
-> Computing predictions with size 192x640
100%|███████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████| 697/697 [01:10<00:00,  9.83it/s]
finished predicting!
-> Evaluating
   Stereo evaluation - disabling median scaling, scaling by 5.4
100%|██████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████| 697/697 [00:03<00:00, 201.34it/s]

   abs_rel |   sq_rel |     rmse | rmse_log |       a1 |       a2 |       a3 | 
&   0.118  &   0.839  &   4.961  &   0.203  &   0.849  &   0.952  &   0.980  \\

Do you have any official results of manydepth trained by MS?