Improving the search accuracy of the VLAD through weighted aggregation of local descriptors

Abstract

We present a novel compact image descriptor, called the Weighted VLAD (wVLAD), which extends the original vector of locally aggregated descriptors (VLAD). The main idea is that the relative importance of local descriptors can be quite different among the local descriptors of an image, depending on the positions from which the descriptors are extracted. Thus, we propose an approach where we assign a weight to each local descriptor of an image, and then compute weighted aggregations of local descriptors. The weights of local descriptors are measured by performing saliency analysis together with an appropriate calibration function. We show, through experiments on publicly available datasets, that our proposed method works better than other existing methods in most image datasets.