ROCAUC Gap between the results in the paper and here

LilitYolyan / CutPaste

Unofficial implementation of Google "CutPaste: Self-Supervised Learning for Anomaly Detection and Localization" in PyTorch

MIT License

114 stars 25 forks source link

ROCAUC Gap between the results in the paper and here #17

Open eliahuhorwitz opened 2 years ago

eliahuhorwitz commented 2 years ago

Hey, Thank you for the implementation! Do you know why is there such a large gap between the results from the paper and the ones you get in this repo? Also, is the localization code working?

Thanks, Eliahu

AnnaManasyan commented 2 years ago

Hey! Thanks, much appreciated! So the gap could be because of the randomness in augmentations also we used smaller batch size not 64 or 96 as stated in the paper due to memory issues we were having. Unfortunately there is a bug in the localization, we are looking into it.

joestrummer82 commented 2 years ago

Hi, could you please comment on the fact why the localization code isn't working currently? Thanks in Advance