the test of image_level in other methods

[giyhub1lili]

Thank you very much for your contribution. May I ask how the model for image forgery localization in table 4 of the paper achieves image-level detection? For example, the iml_vit model, I see that the original paper can only achieve pixel-level localization and cannot achieve image-level detection. )

[iamwangyabin]

Thanks for your question. It is quite straightforward. Similar to previous works, for models designed only for localization, we apply a simple detector: If the model predicts that any pixels in the image are forged, we classify the entire image as fake.

[giyhub1lili]

Thanks for your question. It is quite straightforward. Similar to previous works, for models designed only for localization, we apply a simple detector: If the model predicts that any pixels in the image are forged, we classify the entire image as fake.

Thank you for your response. Do you mean that as long as any pixel is predicted to be counterfeit, the entire image is considered fake? If so, do we need to train a separate decoder head, or can we just add a simple judgment based on the original? I look forward to your reply.

[iamwangyabin]

Yes, that's correct. And no, you don't need to train anything new. You just add a simple judgment based on the original model's output mask. After getting the probability mask, you can threshold it (e.g., at 0.5) to get a binary mask. If the sum of the binary mask is greater than zero, the image is fake. It's a post-processing step.