Regional attention based deep feature for image retrieval

Abstract

Many approaches using Convolutional Neural Network (CNN) for efficient image retrieval have concentrated on feature aggregation rather than feature embedding over recent years, since convolutional features have been found to be reasonably discriminative. Nonetheless, we found that a well-known region-based feature aggregation method, R-MAC, for image retrieval is suffered from the background clutter and varying importance of regions. In this work, we tackle these problems with a simple and effective, context-aware regional attention network that weights an attentive score of a region considering global attentiveness. We conduct various experiments on well-known retrieval datasets, and confirm that our method does not only improve the R-MAC baseline significantly, but also present new state-of-the-art results in the category of ``pre-trained single-pass''. Furthermore, we show that our method shows higher accuracy improvement combined over prior methods, when combined with the query expansion method. These results are attributed by our novel regional-attention network integrated with R-MAC.