Combining Descriptors Obtained through Different Sampling Techniques in Image Retrieval

Abstract

Content-based image retrieval is an important area of research in Multimedia, since it is linked to numerous image applications. Few works in the field have used differently sampled descriptors jointly to improve accuracy, but most of the time the improvement is attributed to other factors. In this paper, firstly we show that a couple descriptor sets can contribute complementary information if each set is sampled differently from the image. To complement the previous fact, we present a simple technique to combine differently sampled descriptors using Fisher Vectors achieving a significant performance improvement at the expense of a small fixed computational cost. Experiments indicate that these observations remain true in scenarios with large image databases.