Few-shot fine-grained classification with Spatial Attentive Comparison

Abstract

The main goal of this paper is to propose a novel model, named Spatial Attentive Comparison Network (SACN), which is used to address a problem, termed few-shot fine-grained recognition (FSFG). FSFG is to recognize fine-grained examples with only a few samples, which is challenging for deep neural networks. SACN is made up of three modules, namely feature extraction module, selective-comparison similarity module (SCSM), and classification module: feature extraction module extracts the distinctive information into feature maps, SCSM is used to fuse the features of support set with those of the query set based on selective comparison. Considering the noisy background and tiny differences between different categories, we apply SCSM to fuse these features by arranging different weights pixel by pixel, and all these weights are learned automatically. Moreover, we apply pyramid structure to enrich the features. By conducting comprehensive experiments on three fine-grained datasets, namely CUB-200-2011 (CUB Birds), Stanford Dogs Dataset, and Stanford Cars Dataset, we demonstrate that the proposed method achieves superior performance over the competing baselines.