Hello everyone I would like to request some information on best practices for training a neural network for image segmentation.
I am conflicted between two approaches lets assume i have patches of 256x256 and my foreground is centered on these patches. Approach one is applying a center crop of lets say 128x128 and do an image translation lets say of 10% which will introduce an artificial border. Approach two is the same but reversed apply the translation first and afterwards perform a center crop.
I am conflicted with these approaches because introducing a border that can vary from image to image seems that it could penalize the borders and it looses partial spatial context despite being faster while second approach is slower but still provides more context rather than the artificial border. Also if first approach does still perform nicely having patches of 256x256 will just be a memory drawback since these can be cropped to smaller.
If anyone could either provide me some help or recommend an article I would be very appreciative.