A family of contextual measures of similarity between distributions with application to image retrieval

Abstract

We introduce a novel family of contextual measures of similarity between distributions: the similarity between two distributions q and p is measured in the context of a third distribution u. In our framework any traditional measure of similarity / dissimilarity has its contextual counterpart. We show that for two important families of divergences (Bregman and Csiszár), the contextual similarity computation consists in solving a convex optimization problem. We focus on the case of multinomials and explain how to compute in practice the similarity for several well-known measures.These contextual measures are then applied to the image retrieval problem. In such a case, the context u is estimated from the neighbors of a query q. One of the main benefits of our approach lies in the fact that using different contexts, and especially contexts at multiple scales (i.e. broad and narrow contexts), provides different views on the same problem. Combining the different views can improve retrieval accuracy. We will show on two very different datasets (one of photographs, the other of document images) that the proposed measures have a relatively small positive impact on macro Average Precision (which measures purely ranking) and a large positive impact on micro Average Precision (which measures both ranking and consistency of the scores across multiple queries). * Yan Liu is a Ph.D. student in the Laboratoire d'Informatique en Image et Systmes d'Information (LIRIS) at the Ecole Centrale de Lyon (ECL).

Exploiting contextual spaces for image re-ranking and rank aggregation

“…Several methods have been proposed for performing re-ranking tasks on various information retrieval systems [2,8,17,[19][20][21]25]. In general, these post-processing methods take an initial ranking and consider some information (relationship among items, user preferences, or other rankings) for improving the effectiveness of ranking.…”

“…In the CBIR scenario, the concept of "contextual information" has been used for designing methods that take information about relationships among images for re-ranking. In [19], the notion of context refers to nearest neighbors of a query, with a similarity measure built for ranked lists. An extension of this approach was proposed in [21].…”

Extracting Structures in Image Collections for Object Recognition

“…We consider the contextual measure, proposed in [29] for the image retrieval task. This context measure is applied for L2 5 to our problem.…”

“…We do not only look at the images themselves, but also at the surrounding images. This is the intuition behind the contextual measure [29], that computes the distance from a first descriptor p to another descriptor q in the context of u using:…”

Combining re-ranking and rank aggregation methods for image retrieval

“…Therefore, the main objective of these methods consists in exploiting the contextual information aiming at improving the distance measures. Significant effectiveness gains have been obtained, considering various algorithms and techniques [21,22,32,34,36,38,41,43,49,62,63].…”