Fine-Grained Few-Shot Classification with Part Matching

Abstract

In this paper, we describe a parts-based approach tailored for fine-grained, few-shot classification, particularly for scenes where the parts distribution is more significant than the broader visual characteristics. By focusing on part-level representations within scenes, our method provides robust classification with limited examples. Our approach, Simple Matching Parts Learner (SMPL), leverages off-the-shelf components in a straightforward manner to optimize few-shot classification using a meta-training phase. We demonstrate the performance of this approach on existing few-shot benchmarks. Additionally, we repurpose an existing fine-grained dataset with higher class diversity and variability than the standard benchmarks for the few-shot setting. SMPL not only achieves state-of-the-art few-shot classification performance, but at a much lower computational cost than compared to the other methods. Code at https://github.com/vidarlab/smpl-fsl.