Modeling the Temporality of Saliency

Abstract

Dynamic cues have until recently been usually considered as a simple extension of the static saliency, usually in the form of optic flow between two frames. The evolution of stimuli over a period longer than two frames has been largely ignored in saliency research. We argue that considering temporal evolution of trajectory even for a relatively short period can significantly extend the kind of meaningful regions that can be extracted from videos, without resorting to higher-level processes. Our work is a systematic and principled investigation of the temporal aspect of saliency under a dynamic setting. Departing from the majority of works where the dynamic cue is considered as an extension of the static saliency, our work places central importance on temporality. We formulate both intra- and inter-trajectory saliency to measure relationships within and between trajectories respectively. Our inter-trajectory saliency formulation also represents the first attempt among computational saliency works to look beyond the immediate neighborhood in space and time, utilizing the perceptual organization rule of common fate (temporal synchrony) to make a group of trajectories stand out from the rest. At the technical level, our use of the superpixel trajectory representation captures the detailed dynamics of superpixels as they progress in time. This allows us to better measure changes such as sudden movement or onset compared to other representations. Experimental results show that our method achieves state-of-the-art performance both quantitatively and qualitatively.