Object retrieval with image graph traversal-based re-ranking

• Our method is a baseline approach based re-ranking method.
• We explore the relation of images in the image level instead of the feature level.
• The proposed image attributes describe the relation between the dataset images and the query from different perspectives.
• The attribute similarity measure and the re-ranking procedure bring much improvements for the baseline retrieval results.
• Our method is also compatible for the multiple-query retrieval with a simple extension.

The topic of this paper is the retrieval of a particular object. A graph traversal-based re-ranking framework for the baseline bag-of-words (BOW) approach is proposed. For an image, we consider not only its similarity with the query image, but also the relationship between other dataset images. We integrate these information as image attributes via an extended image graph and propose a graph traversal algorithm to efficiently obtain their values. By comprehensively considering these attributes, we propose an attribute similarity measure for re-ranking, which brings much performance improvement. We further use our method for the multiple-query retrieval with a simple extension of the virtual query. The experimental results show that our method significantly improve the baseline approach and achieves competitive performance compared with the other state-of-the-art methods. Additionally, our re-ranking method requires only a little extra memory space and time costs.