Skip to content
New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Clarification about CutMix #32

Open
ramanathan831 opened this issue Jan 8, 2021 · 3 comments
Open

Clarification about CutMix #32

ramanathan831 opened this issue Jan 8, 2021 · 3 comments

Comments

@ramanathan831
Copy link

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.

@hellbell
Copy link
Collaborator

hellbell commented Jan 10, 2021

Or is the above mentioned scenario is purely an example? and you replace portions of images randomly?

Yes it is done randomly.

@ramanathan831
Copy link
Author

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?

@hellbell
Copy link
Collaborator

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

Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment
Labels
None yet
Projects
None yet
Development

No branches or pull requests

2 participants