Predicting infant gaze patterns during video viewing using spatiotemporal salience models

Infants’ everyday experiences unfold dynamically across space and time. Although it is well established that visual attention matures significantly across the first postnatal year, relatively little is known about the factors guiding gaze in dynamic contexts. We recorded eye-tracking data from 95 infants (4–12 months) as they viewed brief videos and compared their gaze patterns to the output of three saliency models of increasing complexity: static Graph-Based Visual Saliency (GBVS), dynamic GBVS (incorporating motion and flicker), and TASED-Net, a 3D fully-convolutional neural network that extracts and decodes spatiotemporal features. All models predicted gaze above chance, but the magnitude of prediction differed across models (mean partial correlations: static GBVS, r = .13; dynamic GBVS, r = .18; TASED-Net, r = .29). Dynamic GBVS provided a modest improvement over static GBVS, suggesting limited contributions of motion cues beyond low-level spatial features. TASED-Net strongly outperformed both versions of GBVS. Furthermore, the predictive ability of TASED-Net increased with age, suggesting that older infants’ gaze patterns were increasingly aligned with the abstract spatiotemporal representations learned by the model. These findings establish TASED-Net as a tool for developmental research and suggest that the maturation of gaze control reflects a growing sensitivity to complex spatiotemporal structure.