Developing a Plug-and-Play Meta-Detector for Few-Shot Fine-Grained Image Categorization

Few-shot fine-grained (FSFG) categorization aims to classify new, fine-grained concepts using limited annotated data for each class. It requires a model capable of learning subtle, distinguishable features from similar classes. An intuitive approach is to identify several discriminative parts based on the input image and represent it using these localized parts. It is interpretable since the parts obtained and combined can be seen as the way humans learn about new concepts. Therefore, we propose a straightforward and efficient module for the FSFG task by emphasizing the importance of information from discriminative parts. Specifically, we design a plug-and-play meta-detector based on object detection models, which identifies the positions of distinguishable parts in fine-grained images. With the positions of objects and semantic parts identified, we can extract richer features to enhance few-shot results. Extensive experiments on five fine-grained datasets demonstrate that our approach is applicable to various few-shot backbones and significantly improves accuracy without adding extra parameters. More importantly, compared to other variants on the same base few-shot network, the meta-detector module significantly outperforms them, further demonstrating its effectiveness.