search

MediaBrain-SJTU / RegAD

[ECCV2022 Oral] Registration based Few-Shot Anomaly Detection
MIT License
277 stars 40 forks source link

reproduce the differnet #16

Closed ZhangJiajun1995 closed 2 years ago

ZhangJiajun1995 commented 2 years ago

I use the official source code to reproduce the Differnet in your papers. But my experimental result is lower than yours. I don't konw if you made changes to the details of the experimental settings. If it's convenient, can you share the code about this part?

My email is 1434416800@qq.com

Thanks a lot.

chaoqinhuang commented 2 years ago

We use the official source code in DifferNet to reproduce the results. For fairness of comparison, we extend DifferNet to DifferNet+, where a pre-training procedure is added, i.e., data from multiple categories are used to pre-train the normalizing flow-based model for DifferNet. Then the model is finetuned using its original method with few-shot data from the target category. All you need to do is to modify the dataloader and add a pre-training procedure using its original training codes.