Squeezed Bilinear Pooling for Fine-Grained Visual Categorization

Abstract

In this paper, we propose a supervised selection based method to decrease both the computation and the feature dimension of the original bilinear pooling. Different from currently existing compressed second-order pooling methods, the proposed selection method is matrix normalization applicable. Moreover, by extracting the selected highly semantic feature channels, we proposed the Fisher- Recurrent-Attention structure and achieved state-of-the-art fine-grained classification results among the VGG-16 based models.