Open JohnnieXDU opened 6 years ago
@JohnnieXDU For example, given an 300500 image, first resize it to 448 746 (keep ratio), then random crop (in training) 448*448 patches to be the input of the classification CNN, with the "mirror" parameter being "true". By the way, I think a large "dropout ratio" is important for a promising accuracy.
Hello, can you tell us some details about data augmentation?
thanks.