A deep learning saliency model for exploring viewers' dwell-time distributions over Areas Of Interest on webcam-based eye-tracking data

Visual saliency is a common computational method to detect attention-drawing regions in images, abiding by top-down and bottom-up processes of visual attention. Computer vision algorithms generate saliency maps, which often undergo a validation step in eye-tracking sessions with human participants in controlled labs. However, due to the covid-19 pandemic, experimental sessions have been difficult to roll out. Thus, new webcam-based tools, powered by the developments in machine learning, come into play to help track down onscreen eye movements. Claimed error rates of recent webcam eye trackers can be as low as 1.05°, comparable to sophisticated infrared-based eye-trackers, opening new paths to explore. Using webcams allows reaching a broader participant pool and collecting data over different experiments (e.g., free viewing or task-driven). In our work, we collect webcam eye-tracking data over a collection of images with 2-4 salient objects against a homogenous background. Objects within the images represent our AOIs (areas of interest). We have two main goals: a) Check how eye movements vary on AOIs across all spatial permutations of the same AOI in a given image; b) Extract correlations for a given image containing N objects between viewers’ eye movement dwell times over the N AOIs and the corresponding AOIs saliency maps. We will show relationships between viewers’ dwell time over each AOI throughout all factorial N spatial permutations and variance of AOIs’ salient pixels. Based on this relationship, eventually, object-oriented saliency models can be used to predict dwell-time distributions over AOIs for a given image.