I am training a 'specialist network' to reconstruct images of an object using a Variational Autoencoder (VAE). The training set (~15000 images) is of a single object in multiple poses.
I also want this network to classify the object identity as being either the object it was trained on (class 1) VS anything else (class 0).
Is this possible using Binary Cross Entropy (BCE) loss and ONLY images of class 1? Or will it need negative examples (class 0)? If it does need negative examples, will these need to be exhaustive to ensure accuracy of classifying class 0? e.g. previously unseen images could yield class 1 if not included as class 0 in training.
Is there another method to achieve this other than BCE?
Best Answer
Because you only have data about 1 class, standard classification approaches are not applicable. Ideally, you could just collect data for the other classes, and then employ a standard approach, but that's not always feasible.
The term of art for this task is "one-class classification." Here's a review article from a few years back -- there have been a number of different options proposed.
Khan, S., & Madden, M. (2014). One-class classification: Taxonomy of study and review of techniques. The Knowledge Engineering Review, 29(3), 345-374. doi:10.1017/S026988891300043X