The Effect of Visual Defects in 3D Models on Visual Attention: A Desktop and VR Eye-Tracking Study
Erwan David1, Sonain Jamil1, Julien Pettré2, Sébastien Iksal1

1 Le Mans Université, LIUM (FR)
2 Inria, University Rennes, IRISA (FR)

Virtual reality simulations are transforming education and training by providing controlled, immersive environments for skills that are costly, hazardous, or complex to teach through traditional methods. Central to simulation effectiveness is fidelity: the degree to which virtual environments accurately replicate real-world objects and phenomena. Low fidelity may disrupt learner attention, compromising educational outcomes, yet the field lacks theoretical frameworks and tools to measure and mitigate fidelity-related distractions. The study I will present addresses this gap by investigating visual attention as an objective indicator of simulation fidelity. Specifically, we examine how visual defects in 3D-modelled objects distract viewers and whether immersiveness modulates this effect. We conducted a 3D-mesh viewing protocol with eye tracking across two viewing conditions: desktop display and head-mounted VR. Participants viewed 3D objects containing controlled visual defects (mesh simplification, self-intersection, smoothing, and semantic) while gaze behaviour was recorded as a proxy for visual attention allocation. Each participant was presented with half of the objects in their unmodified state to mitigate an expectation bias. Gaze data were analysed using region-of-interest metrics (dwell time, time to first fixation) and saccadic measures (saccade angles, scanpath patterns). We compared between defect conditions to characterise how they may capture and impact attention patterns across viewing contexts. Results characterise attentional capture by defect type and viewing condition, revealing systematic desktop-VR differences. From these data, we propose a quantitative ranking of the types of defect. This work is a step towards a predictive model of perceived realism in 3D simulations. Such models have broad applications across industries using or relying on simulated stimuli, from educational VR to entertainment and industrial training. We hope that our findings will provide clear and usable metrics, to quantify simulation fidelity and identify visual defects, that will help improve learner attention and optimise educational impact.