Hedging static saliency models to predict dynamic saliency

Yasin Kavak, Erkut Erdem, Aykut Erdem

Abstract: Highlights•We propose a novel weakly supervised dynamic saliency model called HedgeSal.•Our model depends on a decision-theoretic online learning algorithm.•We process appearance and motion streams via pre-trained deep static saliency models.•Combining their decisions in an adaptive manner yields state-of-the-art results.