Test result using ur weight

[EricZhou0221]

Hello, I am a first-year master's student from ShenZhen University, I have conducted testing using your provided weight parameters on seven basic datasets, but the results do not align with those stated in the paper. My approach involved calculating F1 and AUC values for each image (due to the binary nature of the model's output mask, AUC is calculated for a single point), followed by computing the mean. All the results I obtained were mean values.
I would like to inquire whether the weight parameters you provided are pre-trained or obtained through benchmark training.

[Knightzjz]

Hi Zhou, sorry for the late reply. I am on my leave for a few days.
I checked the pre-trained weights on colab and found I might have uploaded an incorrect version due to my bad naming habit. I will further check the pre-train weights issue when I head back to the campus with my desktop aside. Thanks for pointing out this issue for me.
BTW, if you are following our research on IML, you may check this answer. According to my arrangement, the entire training code for our latest benchmark model IML-ViT is about to be released. I suggest you follow IML-ViT's structure since we recently confirmed that the ViT architecture is much more powerful than CNNs in IML or anti-AIGC tasks and we will totally open-source this model. Also, IML-ViT is now accepted by AAAI24.
Again, thanks for your careful test, I will be correcting the pre-training weights A.S.A.P.

[EricZhou0221]

Hi Zhou, No worries about the delay, and thanks for getting back to me. I appreciate your thorough investigation into the pre-trained weights issue. It's completely understandable, and I look forward to the corrected version. I will continue to follow and learn from the relevant research in IML. I am about to read and study the paper and associated code you released about IML-ViT. Thank you for your guidance and CONGRATULATIONS on your paper being accepted by AAAI.