Few‐shot classification using Gaussianisation prototypical classifier

Few-shot classification (FSC) aims at classifying query samples into correct classes given only a few labelled samples. Prototypical Classifier (PC) can be chosen to an ideal classifier for settling this problem, as it has good properties of low-capacity and parameter-free. However, the mean-based prototypes suffer from issue deviating its ground-truth centre. In order solve such problem prototype bias, Gaussianisation (GPC) is proposed, which kind one-step rectification method. Specifically, authors first perform operation over feature extracted backbone network so that features fit particular Gaussian distribution. Second, use base class prior information employs Maximum Posteriori estimation method obtain reliable each novel class. Finally, sample classified nearest with non-parametric classifiers. Extensive experiments have been conducted on multiple FSC benchmarks. Comparative results also demonstrate authors’ superior existing state-of-the-art methods.