Top down saliency estimation via superpixel-based discriminative dictionaries

Abstract

Predicting where humans look in images has gained significant popularity in recent years. In this work, we present a novel method for learning top-down visual saliency, which is well-suited to locate objects of interest in complex scenes. During training, we jointly learn a superpixel based class-specific dictionary and a Conditional Random Field (CRF). While using such a discriminative dictionary helps to distinguish target objects from the background, performing the computations at the superpixel level allows us to improve accuracy of object localizations. Experimental results on the Graz-02 and PASCAL VOC 2007 datasets show that the proposed approach is able to achieve stateof- the-art results and provides much better saliency maps.