What has been missed for predicting human attention in viewing driving clips?

Recent research progress on the topic of human visual attention allocation in scene perception and its simulation is based mainly on studies with static images. However, natural vision requires us to extract visual information that constantly changes due to egocentric movements or dynamics of the world. It is unclear to what extent spatio-temporal regularity, an inherent regularity in dynamic vision, affects human gaze distribution and saliency computation in visual attention models. In this free-viewing eye-tracking study we manipulated the spatio-temporal regularity of traffic videos by presenting them in normal video sequence, reversed video sequence, normal frame sequence, and randomised frame sequence. The recorded human gaze allocation was then used as the 'ground truth' to examine the predictive ability of a number of state-of-the-art visual attention models. The analysis revealed high inter-observer agreement across individual human observers, but all the tested attention models performed significantly worse than humans. The inferior predictability of the models was evident from indistinguishable gaze prediction irrespective of stimuli presentation sequence, and weak central fixation bias. Our findings suggest that a realistic visual attention model for the processing of dynamic scenes should incorporate human visual sensitivity with spatio-temporal regularity and central fixation bias.