Efficient graph spanning structures for large database image retrieval

In this paper we propose a novel method to improve the performance of image retrieval at VLAD descriptor level. The system performs image re-ranking based on relational graphs and neighborhood relations of the top-k candidate results. The technique is able to treat differently various parts of the graph spanning structures by adaptively modifying the similarity score between images. Because most of the processing is performed offline, our algorithm does not influence the retrieval time. By dealing with uneven distribution of images in the dataset, the method is effective and increases the accuracy without relying on low-level information or on the geometrical verification of the considered features.