Learning Representation of Multi-Scale Object for Fine-Grained Image Retrieval

Abstract

Extracting discriminative local features has attracted many research focus in fine-grained image retrieval task. With attention mechanism and softmax-like loss functions, deep neural networks could locate and learn the representation of the most discriminative region of objects, however, which also makes other non-most discriminative regions be ignored to some extent. In our work, to extract more local features, we propose a method that could proposes multiple discriminative regions on different scales, which could provide more refined local and multi-sacle representation for fine-grained image retrieval. Experimental results show that our proposed method achieves excellent performance on two benchmark fine-grained datasets, which demonstrates its effectiveness.