Integrating Action-aware Features for Saliency Prediction via Weakly Supervised Learning

Deep learning has been widely studied for saliency prediction. Despite the great performance improvement introduced by deep saliency models, some high-level concepts that contribute to the saliency prediction, such as text, objects of gaze and action, locations of motion, and expected locations of people, have not been explicitly considered. This paper investigates the objects of action and motion, and proposes to use action-aware features to compensate deep saliency models. The action-aware features are generated via weakly supervised learning using an extra action classification network trained with existing image based action datasets. Then a feature fusion module is developed to integrate the action-aware features for saliency prediction. Experiments show that the proposed saliency model with the action-aware features achieves better performance on three public benchmark datasets. More experiments are further conducted to analyze the effectiveness of the action-aware features in saliency prediction. To the best of our knowledge, this study is the first attempt on explicitly integrating objects of action and motion concept into deep saliency models.