Improving CNN classifiers by estimating test-time priors

Milan Sulc, Jiri Matas

The problem of different training and test set class priors is addressed in the context of CNN classifiers. We compare two different approaches to estimating the new priors: an exi…