clovaai
Clarification about CutMix
I saw examples where head of a dog is cut and overlayed with head of a cat. How do identify these specific portions of the body - or how do identify these important portions of the two objects. Or is the above mentioned scenario is purely an example? and you replace portions of images randomly?
I hope I am clear in my question, else I shall rephrase it.
Or is the above mentioned scenario is purely an example? and you replace portions of images randomly?
Yes it is done randomly.
So there are chances that the main features (like head) of a dog won't be replaced right? Which in turn means paying attention to non significant parts of the object won't happen which is an important reason for improvement in scores right.
If this important thing doesn't happen how can we explain the effectiveness?
So there are chances that the main features (like head) of a dog won't be replaced right? Which in turn means paying attention to non significant parts of the object won't happen which is an important reason for improvement in scores right.
If this important thing doesn't happen how can we explain the effectiveness?
Paying attention to non-significant parts of the object will happen since CutMix does at random regions (e.g., there can be only legs of a cat and a dog when mixing the cat and dog images). Our hypothesis is that learning of non-significant regions (e.g., backgrounds) is important as well to get improved image classification performance.
Or, it would be better to check this ICML paper https://icml.cc/Conferences/2020/ScheduleMultitrack?event=6827