Recurrent Refinement for Visual Saliency Estimation in Surveillance Scenarios

In recent years, many different proposals for visual saliency computation have been put forth, that generally frame the determination of visual saliency as a measure of local feature contrast. There is however, a paucity of approaches that take into account more global holistic elements of the scene. In this paper, we propose a novel mechanism that augments the visual representation used to compute saliency. Inspired by research into biological vision, this strategy is one based on the role of recurrent computation in a visual processing hierarchy. Unlike existing approaches, the proposed model provides a manner of refining local saliency based computation based on the more global composition of a scene that is independent of semantic labeling or viewpoint. The results presented demonstrate that a fast recurrent mechanism significantly augments the determination of salient regions of interest as compared with a purely feed forward visual saliency architecture. This demonstration is applied to the problem of detecting targets of interest in various surveillance scenarios.