A Coarse-to-Fine Human Visual Focus Estimation for ASD Toddlers in Early Screening

Human visual focus is a vital feature to uncover subjects' underlying cognitive processes. To predict the subject's visual focus, existing deep learning methods learn to combine the head orientation, location, and scene content for estimating the visual focal point. However, these methods mainly face three problems: the visual focal point prediction solely depends on learned spatial distribution heatmaps, the reasoning process in post-processing is non-learnable, and the learning of gaze salience representation could utilize more prior knowledge. Therefore, we propose a coarse-to-fine human visual focus estimation method to address these problems, for improving estimation performance. To begin with, we introduce a coarse-to-fine regression module, in which the coarse branch aims to estimate the subject's possible attention area while the fine branch directly outputs the estimated visual focal point position, thus avoiding sequential reasoning and making visual focal point estimation is totally learnable. Furthermore, the human visual field prior is used to guide the learning of gaze salience for better encoding target-related representation. Extensive experimental results demonstrate that our method outperforms existing state-of-the-art methods on self-collected ASD-attention datasets.