generalization problem

[kkpssr]

first, thanks for the work and i notice some problem. the final predict matrix is simply divided into [0,1,2,3,4],its that mean the detector model is only suitable for camera model in [0,1,2,3,4] or it is not support tamper from unseen camera model? but in real world, its impossible to get datasets included all camera model . and in the type of copy move, the forgery one and splice one is from the same camera model ,how to address this situation ? and is there some information for the performance in different datasets,i have tried some work in this field ,and most of them shows serious overfitting.