Abstract
Eye tracking is commonly used in visual neuroscience and cognitive science to answer related questions such as visual attention and decision making. We envision that bigger eye-tracking data can advance the understanding of these questions due to the inherently complex nature of both the stimuli and the human cognitive process. The scale of current eye-tracking experiments, however, are limited as it requires a customized device to track gazes accurately. Low-cost data collection with general-purpose webcams is not yet accurate enough, especially in uncontrolled environments. We present a new method to allow the collection of large-scale attentional data using a gaze-contingent multi-resolutional mechanism. Subjects can move the mouse to direct the high-resolutional fovea to where they find interesting in the image stimuli. The stimuli encoded the visual acuity drop-off as a function of retinal eccentricity. The mouse-contingent paradigm motivated mouse movements, to reveal interesting objects in the periphery with high resolution, similarly as humans shift their gazes to bring objects-of-interest to the fovea. The mouse trajectories from multiple subjects were aggregated to indicate where people look most in the images. The new paradigm allowed using a general-purpose mouse instead of an expensive eye tracker to record viewing behaviours, thus enabling large-scale attentional data collection. We propose a crowdsoucing mechanism to collect mouse-tracking data through Amazon Mechanic Turk, and a proof-of-concept dataset of 60 subjects “free-viewing” 10,000 images from the recent Microsoft COCO dataset. The crowdsourcing allowed us to easily collect and compare various data with different top-down instructions. Our results suggested that the large-scale mouse-tracking data were much more similar to the eye-tracking data than those predicted from the state-of-the-art computational saliency models. They were also shown to be a good source as ground truth for the evaluation of saliency algorithms.
Meeting abstract presented at VSS 2015