Objects in focus: How object spatial probability underscores eye movement patterns

In human vision, objects and scenes are fundamentally interconnected elements that play a critical role in shaping our perception and understanding of the visual world. In natural viewing conditions, objects and scenes are seldom encountered in isolation. The visual system has evolved to process the statistical probability of different spatial arrangements, with contextual information offering key insights into object functionality and relevance. One way to elucidate the human visual perceptive process is to track and predict how humans visually explore images through eye movements. Other datasets of eye movements have been shared openly online. However, existing datasets are limited by the range of camera perspectives represented (close-up viewpoints of objects or faces vs. wide-angle perspectives of real-world scenes), and the number of objects segmented within the images. To address this, we developed the Objects in Focus dataset. We semantically segmented 2800 unique objects across the 100 real-world scenes and collected measures of each object's size, eccentricity, semantic significance, and visual saliency, with an average of 28.7 segmented objects in each scene image. We also include analyses comparing the distributions of objects and fixations across scene images and discuss how image size and perspective influence attention dispersion. Finally, we relate our findings considering attention prediction networks and discuss how formal definitions of spatial arrangement probabilities can be used to improve different computer vision models.