Abstract

Top-down saliency detection is a knowledge-drivensearch task. While some previous methods aim to learn this“knowledge” from category-specific data, others transferexisting annotations in a large dataset through appearancematching. In contrast, we propose in this paper a locate-by-exemplar strategy. This approach is challenging, as weonly use a few exemplars (up to 4) and the appearancesamong the query object and the exemplars can be very dif-ferent. To address it, we design a two-stage deep model tolearn the intra-class association between the exemplars andquery objects. The first stage is for learning object-to-objectassociation, and the second stage is to learn backgrounddiscrimination. Extensive experimental evaluations showthat the proposed method outperforms different baselinesand the category-specific models. In addition, we explorethe influence of exemplar properties, in terms of exemplarnumber and quality. Furthermore, we show that the learnedmodel is a universal model and offers great generalization to unseen objects.