Training image size

[ghost]

I'd like to train on Kitti dataset, which is of size 1242 x 375. Would this work with the current pool size?

Update: Tried to train with the default parameters from train.py, but after 2000 iterations the loss is ~2.