Motion statistics at the saccade landing point: attentional capture by spatiotemporal features in a gaze-contingent reference
Anna Belardinelli, Andrea Carbone · Proceedings of SPIE, the International Society for Optical Engineering/Proceedings of SPIE · 2012
Motion is known to play a fundamental role in attentional capture, still it is not always included in computational models of visual attention. A wealth of literature in the past years has investigated natural image statistics at the centre of gaze to assess static low-level features accounting for fixation capture on images. A motion counterpart describing which features trigger saccades on dynamic scenes has been less looked into, whereas it would provide significant insight on the visuomotor behaviour when attending to events instead of less realistic still images. Such knowledge would be paramount to devise active vision systems that can spot interesting or malicious activities and disregard less relevant patterns. In this paper, we present an analysis of spatiotemporal features at the future centre of gaze to extract possible regularities in the fixation distribution to contrast with the feature distribution of non-fixated points. A substantial novelty in the methodology is the evaluation of the features in a gaze-contingent reference. Each video sequence fragment is indeed foveated with respect to the current fixation, while features are collected at the next saccade landing point. This allows us to estimate covertly selected motion cues in a retinotopic fashion. We consider video sequences and eye-tracking data from a recent state-of-the art dataset and test a bottom-up motion saliency measure against human performance. Obtained results can be used to further tune saliency computational models and to learn to predict human fixations on video sequences or generate meaningful shifts of active sensors in real world scenarios.