Objects guide human gaze behavior in dynamic real-world scenes

Roth, Nicolas; Rolfs, Martin; Hellwich, Olaf; Obermayer, Klaus · 2023 · OpenAlex-citations

DOI: 10.1371/journal.pcbi.1011512

archive: archived pipeline: cataloged verified

Get this paper ↗ (DOI — opens at the source; we link to it, we don't host it)

Summary

This study addresses the challenge of modeling human gaze behavior in dynamic, real-world scenes, specifically investigating whether visual attention is guided by spatial saliency or by object-based units. Historically, eye movement models relied on space-based attention driven by low-level visual features. However, growing evidence suggests that objects serve as fundamental attentional units. Existing computational models are largely limited to static images or predict only average fixation distributions, failing to capture the temporal dynamics (saccade timing) and smooth pursuit behaviors inherent in viewing moving scenes. To bridge this gap, the authors developed ScanDy, a modular, mechanistic computational framework designed to simulate realistic scanpaths in dynamic environments. The ScanDy framework integrates psychophysically grounded mechanisms for visual attention and saccadic decision-making. It models gaze behavior through a sequential decision process using a drift-diffusion model, where evidence accumulates for potential saccade targets until a threshold is reached. The framework accounts for visual sensitivity, scene features, inhibition of return, and smooth pursuit eye movements. To systematically evaluate the role of objects, the authors implemented five distinct model variants within this framework: two purely spatial models (one using low-level saliency, one using high-level saliency), two object-based models (one incorporating low-level saliency for object prioritization, one using only center bias), and a mixed model combining object-based attention with space-based inhibition of return. Model parameters were optimized using evolutionary algorithms to reproduce human saccade amplitude and fixation duration distributions from the VidCom dataset, which contains free-viewing eye-tracking data on videos. The results demonstrate that models incorporating object-based attention and selection significantly outperform purely spatial models. Specifically, the object-based model that uses saliency information to prioritize between objects for saccadic selection (the O.ll model) produced scanpath statistics with the highest similarity to human data. This model accurately replicated spatial and temporal fixation behaviors, including the proportions of fixations dedicated to detecting, inspecting, and returning to objects, as well as exploring the background. In contrast, spatial models failed to capture these nuanced exploration patterns. The findings indicate that object-level attentional units play a critical role in guiding attentional processing in dynamic scenes, more so than spatial saliency alone. The significance of this work lies in providing a robust, interpretable framework for studying visual attention in ecologically valid, dynamic contexts. By demonstrating that object-based mechanisms are essential for accurately simulating human gaze behavior, the study challenges the dominance of space-based saliency models. The open-source availability of the ScanDy framework encourages further research into the mechanisms of visual attention, offering a tool that bridges the gap between controlled psychophysical experiments and complex real-world viewing behavior. This approach allows for the systematic testing of hypotheses regarding how humans allocate overt attention in changing environments.

Provenance

The full processing record for this entry. Every stage of this paper's journey through the pipeline is logged — what ran, with which tool and model, how many attempts it took, and when it last completed.

Summary generated by qwen3.6-27b-prismaquant on 2026-06-25; verification: verified.

Topics

Ranked by relevance to this paper. Hover a topic for its definition.

• eye movements scanning
• gaze based attention detection
• attention allocation