#random-cropping

Question 1: Single choice

You add random cropping as data augmentation when training an object detector. What has to be handled that does not come up in image classification?

Boxes must be shifted and clipped to the crop, and objects left mostly outside it dropped from the labels1
The crop must be resized back to the original size, which keeps every box coordinate valid2
Nothing extra, because the labels of an image stay valid after any crop of that image3
Only crops that contain no objects may be used, so that no bounding box is ever cut4