OBJECTIVE PREDICTION OF VISUAL SALIENCY MAPS IN EGOCENTRIC VIDEOS FOR CONTENT-ACTION INTERPRETATION