Few-Shot Prompt Data Anomaly Filtering in 'dataset/dataset.py'

[cjm-jamie]

Hi,  

I'm reviewing 'dataset/dataset.py' for the few-shot setting and have a quick question.  
My understanding is that few-shot prompts should use normal (non-anomalous) images. However, at lines 227 and 240, data_tmp (used for prompts) seems to be assigned directly from meta_train_info[cls_name] without explicitly filtering for item['anomaly'] == 0.  

Shouldn't data_tmp be filtered to ensure only normal images are used, e.g., data_tmp = [item for item in data_tmp if item['anomaly'] == 0]?  

Am I misunderstanding the intended behavior here?  

Thanks for your great work!

[gaobb]

hi,

Thank you for pointing out this issue. Your understanding is correct. However, there is also no problem with our original code not filtering out anormaly images.

1. Generally, in standard anomaly detection datasets (e.g., MVTec and Visa), all training images are assumed to be normal. Therefore, the logic for filtering out anomaly images is not necessary;

2. In special cases, assuming the training dataset (such as Kvasir and Endo) does not contain normal images, we still default to randomly sampling few-shot anomaly images from the training set as visual prompts.

3. If the training dataset contains a mix of normal and anormaly images, filtering out anormaly images as you suggested would be more reasonable.

If you have any further questions, please feel free to discuss.