Revisit augmentation strategy

[dshean]

I'm not sure random horizontal and vertical flips necessarily make sense given physical constraints for solar illumination, shadows, and relief displacement in the orthoimages (esp relevant for the off-nadir image).
Imagine a single tree casting a shadow. With H and V flips, we learn 4 possible "shadow directions" and a subset of convolution kernels will be most relevant. But the input images for inference are always "north-up" and the directions of shadows will only be in one direction.
I'm not sure how this will impact model transfer for new images with different solar illumination or view geometry.
Maybe only H flips make sense here. Can revisit with geometry handling @ayush12gupta