search

caoyunkang / CDO

[TII 2023] Collaborative Discrepancy Optimization for Reliable Image Anomaly Localization
MIT License
63 stars 7 forks source link

MEmodel update #11

Open zq130320339 opened 7 months ago

zq130320339 commented 7 months ago

Hi, have you tested the effect of MEmdel doing gradient update vs. no update, I tested that the difference is not that much

caoyunkang commented 7 months ago

Hi, I have not tried this. However, it may be reasonable to update it in the first several epochs, so that the ME model is adapted to the target category, mitigating the domain gap. Just be careful that too much training on the ME model may lead to overfitting. Best.

zq130320339 commented 7 months ago

Thanks for the reply, I tried adding some supervised data training, but the test was not very good, some images in the test set that were similar to the training performed poorly.

caoyunkang commented 7 months ago

Can you describe your problem more clearly? Try to use different backbones and the performance may vary.