The methods is just equivalent to augment more data within one epoch? #8
Description
I am really confused by your claim of reducing two orders of magnitudes training epochs. Since you are cropping one image to 200 same-size images, is it equivalent to a very heavy augmentation? Even though you have much fewer epochs, the actual computational cost/GPU hours do not reduce significantly compared to other methods. (Which is also reflected in Table 3). Did i miss something or is it what you actually did?