Changing Teacher PDN for EfficientAD

[WladimirLct]

Hi, I am currently working on an anomaly detection project, and I stumbled across a problem :
The performance of my model (EfficientAD) is quite deceiving : To get a 95% recall, I have ~0.5 False positive rate.
I feel like this might be induced by the Teacher PDN's pretraining, since it learns on 256x256 images whereas my images need to have a higher resolution to maintain their defects visible (base is 1920x1080 and I could reduce to 512x512 or 384x384).
Do you feel like it would help to redo a distillation of a WideResNet on a bigger PDN ?

[abc-125]

Hello, EfficientAD's teacher was pretrained to simulate results of WideResNet101 on 512x512 input, so in a way, it is already 512x512 even if input size is 256x256.

You probably can simply try a higher input size for EfficientAD training, 512x512 or so. You can also try different models, PatchCore or ReverseDistillation, also with higher input size. Be aware that higher input size brings more computational constraints and also can reduce the model's ability to detect large defects.

[WladimirLct]

Hi, thank you very much for your answer.

I have not expressed my problem correctly in my initial question, but for my use case it's infeasible to reduce to 256x256 without losing vital information on defects. I understand what you're saying, but even though the PDN is trained to replicate WideResNet's features on 512x512 images it nonetheless uses downscaled 256x256 images as input and uses an optimized 33x33 receptive field for that resolution.

I did try changing the input size of EfficientAD (e.g. 256, 384 and 512) before opening this topic but performance stayed the same. I feel this is likely due to EfficientAD's way of accepting higher resolution images (only modifying the AE's architecture while keeping the same PDNs).

Since the PDN's architecture remains unchanged, the receptive field increases with the input resolution. Don't you think that this could be the explanation of the ST missing subtle anomalies ?
I'm interested to know your perspective on this.

Originally I wanted to retrain or modify the PDN's architecture to either teach it how to describe bigger patches or reduce receptive field back to a more reasonable size. For now I just went with an EfficientAD ensemble model using overlapping patches of 256x256 as well as implementing a latency mechanism, as my images come from videos to reduce the FP rate.