Open jacoblam3112 opened 1 month ago
Hello !
I also do not have a clear answer to this question. In the original design of RealNet, the reconstruction residuals already contain a wealth of anomaly information. The role of the discriminator is very simple, possibly serving merely a normalization function, with the primary task of anomaly detection being carried out by the reconstruction network.
It would technically be reasonable to remove the reconstruction network and use only the discriminator to differentiate between SDAS images and normal images. If there is a significant performance drop, it may indicate that there is a difference between the synthesis anomaly image and the real anomaly images contained in the dataset, possibly suggesting that the discriminator is more sensitive to this distribution difference? This could also be related to the design of the discriminator.
Thank you for your prompt and detailed explanation.
Hi, thank you for sharing your work. I would like to ask a simple question to help strengthen my understanding. If I directly use the multi-scale backbone features and skip the reconstruction stage and then train the discriminator to identify anomalies based on normal vs SDAS images features, why would this not work as good as using the reconstruction residuals ?