Dual adaptive representation of vector of locally aggregated

Abstract

This paper addresses the problem of large-scale image retrieval. We use the dual adaptive representation of vector of locally aggregated to improve the retrieval efficiency. The vector of locally aggregated (VLAD) aggregates SIFT descriptors and produces a compact representation to improve the search accuracy and memory usage, and the usage of adapted cluster centers of the VLAD enhances the performance further. We first carry out twice adaptation on the cluster centers to optimize the references of the features which are used to calculate the center residuals, and to obtain the vector of an image by jointing the center residuals of each corresponding cluster in the initial retrieval process. We then reduce the dimensionality of the vectors by using PCA, and evaluate the similarities between query image the top N result image by the residual of sparse representation in the re-rank process. Finally, experiments show clearly that our work improves the retrieval accuracy.