Feature hallucination in hypersphere space for few‐shot classification

Few-shot classification (FSC) targeting at classifying unseen classes with few labelled samples is still a challenging task. Recent works show that transfer-learning based approaches are competitive meta-learning ones, which usually pre-train convolutional neural networks (CNN)-based network using cross-entropy (CE) loss and throw away the last layer to post-process novel classes. Hereby, they suffer issue of getting more transferable extractor lacking enough samples. Thus, authors propose algorithm feature hallucination in hypersphere space (FHHS) for FSC. On first stage, (HL), supplies CE supervised contrastive (SC) self-supervised (SSL), SC can map base images onto densely. second generate new their synthetic sampling (SNSB), linearly interpolate between each class prototype its K nearest neighbour prototypes. Comprehensive experiments on multiple popular FSC demonstrate HL enhance performance backbone authors’ method superior existing hallucination-based methods.