HEVC stream saliency extraction: Synergies between FIT and information theory principles

Abstract

The present paper studies the potential synergies between a popular approach in saliency extraction — FIT (feature integration theory) and source coding principles. By combining these two approaches, a new saliency model, extracted directly at the HEVC steam syntax elements level is defined. The experiments confront the new model to the human saliency, captured by eye-tracking devices. They consider a reference corpus representing density fixation maps, two objective criteria, two objective measures and 7 state-of-the-art saliency models (3 acting in pixel domain and 4 acting in compressed domain).