Hi all,
thanks for this nice implementation!
As far as I understand from the WinCLIP paper, the approach should be architecture agnostic, meaning that one can use any image and text embeddings to obtain window-based anomaly scores.
Therefore, I was trying to adapt other architectures (e.g. the ResNet and/or other transformers like SwinTransformer).
However, I think I am not fully understand the masking process and how it can be adapted to other architectures.
Can you give some hints?
Thanks!
Hi all,
thanks for this nice implementation!
As far as I understand from the WinCLIP paper, the approach should be architecture agnostic, meaning that one can use any image and text embeddings to obtain window-based anomaly scores.
Therefore, I was trying to adapt other architectures (e.g. the ResNet and/or other transformers like SwinTransformer).
However, I think I am not fully understand the masking process and how it can be adapted to other architectures.
Can you give some hints?
Thanks!