lorenmt / reco

The implementation of "Bootstrapping Semantic Segmentation with Regional Contrast" [ICLR 2022].
https://shikun.io/projects/regional-contrast
Other
162 stars 25 forks source link

question about Ablative Analysis #15

Closed wwjwy closed 2 years ago

wwjwy commented 2 years ago

Hi, In Section 4.4 Ablative Analysis, the Effect of Active Sampling was verified. Could you please tell me how to conduct (Random Query,Random Key ) and other settings? Thanks!

lorenmt commented 2 years ago

Hello!

That's fairly simple. Using a very large temperature via --temp 100 for example, will roughly give a uniform sampling, i.e. random key sampling. Using a very low weak threshold and a very large strong threshold via --weak_threshold 0.0, --strong_threshold 1.0 will then sample all pixels as query, i.e. random query sampling.

Hope that helps.

wwjwy commented 2 years ago

Thank you very much! can you tell me how to realize Easy Query? other settings i have understand your meaning.

lorenmt commented 2 years ago

Change: https://github.com/lorenmt/reco/blob/265cd2c97b8bf5dda461034f1dd08e4fb75c5c1e/module_list.py#L87 into > strong threshold

wwjwy commented 2 years ago

Change:

https://github.com/lorenmt/reco/blob/265cd2c97b8bf5dda461034f1dd08e4fb75c5c1e/module_list.py#L87

into > strong threshold

Does the strong_threshold should keep the original setting(0.97)?

lorenmt commented 2 years ago

Yes.

wwjwy commented 2 years ago

Thank you very much for your timely help!