A saliency model for goal directed actions

In this paper, we propose a saliency model which can be used to guide eye movements for viewing a goal-directed action video. The model employs top-down and bottom-up saliency components which work purely on contrasts of random pixels in the image. We construct task specific spatio-temporal priors and integrate them into the top-down and bottom-up modules. These priors reduce the search space for the target object, thereby automatically suppressing those image regions that are task irrelevant. For the purpose of evaluation we introduce a new goal-directed video database containing 60 sequences of five goal-directed actions performed by the same human demonstrator. The presented results of saliency detection justify the proposed model and demonstrate its advantage over models for task independent saliency detection.