Revisiting Local Descriptor for Improved Few-Shot Classification

Few-shot classification studies the problem of quickly adapting a deep learner to understanding novel classes based on few support images. In this context, recent research efforts have been aimed at designing more and complex classifiers that measure similarities between query images but left importance feature embeddings seldom explored. We show reliance sophisticated is not necessary, simple classifier applied directly improved can instead outperform most leading methods in literature. To end, we present new method, named DCAP, for few-shot classification, which investigate how one improve quality by leveraging Dense Classification Attentive Pooling (DCAP) . Specifically, propose train base with abundant samples solve dense first then meta-train plenty randomly sampled tasks adapt it scenario or test time scenario. During meta-training, suggest pool maps applying attentive pooling widely used global average prepare classification. learns reweight local descriptors, explaining what looking as evidence decision making. Experiments two benchmark datasets proposed method be superior multiple settings while being simpler explainable. Code publicly available https://github.com/Ukeyboard/dcap/