Objective Prediction of Visual Saliency Maps in Egocentric Videos for Content-action Interpretation

Extraction of visual saliency from video is in the focus of intensive research nowadays due to the variety and importance of application areas. In this paper we study the relation between subjective saliency maps, recorded on the basis of gazetracker data in a new upcoming video content: the egocentric video recorded with wearable cameras. On the basis of physiological research and comparing the subjective maps of an Actor performing activities of everyday life and a Viewer who interprets the video after it has been recorded, we identify the temporal shift between these two saliency maps. Using this relation we propose an ”à la carte” prediction of saliency maps of an Actor for the beginning of actions by an objective saliency model we previously developed. All the components of objective saliency: spatial, temporal and central bias are merged in this prediction. The commonly used quality metrics for pixel-based saliency prediction such as Pearson Correlation Coefficient, Normalized Scan Path and Area Under Curve show the good correspondence of predicted maps for Actor and Viewer. This research seems to us promising for content interpretation coming from mobile video recording devices.