I wonder what is the best augmentation strategy. Should I start with no augmentation and just when it is overfitting change the model data to start augmentation? Or should I start with strong augmentation from the beginning?
I am working with segmentation in a problem similar to Carvana (lesson 14). I get better results with weak (almost none) augmentation, weird. Maybe this way it is getting good results with the data but a note generalizable solution.
