Salience-based object prioritization during active viewing of naturalistic scenes in young and older adults

The question of what guides our attention and gaze during naturalistic scene viewing has sparked debate between a 'salience view' (fixations driven by low-level image features) and an 'object view' (fixations driven by objects). This research aims to reconcile these views by comparing the effects of location-based and object-based visual salience on fixation selection in young and older adults (65+ years). Computational models, particularly saliency map algorithms, suggest that visually salient regions attract attention first. These models predict human fixation selection reasonably well. However, the alternative 'object view' proposes that objects are primary units of saccadic selection. Evidence supporting this view includes the observation that fixations within objects are distributed according to a 2D Gaussian distribution, with a mean close to the object's center (Preferred Viewing Location, PVL). This PVL is modulated by factors such as object size, saccade direction, and object category. The debate centers around whether saliency maps predict fixations directly or indirectly, by estimating object locations. Some research suggests that once object locations are known, saliency maps offer little additional predictive power. However, object-based models incorporating the PVL predict fixations equally well as or better than low-level saliency models, especially when scenes are manipulated to dissociate objects from high-salience regions. A critical question is how observers prioritize among multiple objects. This study investigates this question while also exploring age-related differences in eye guidance. Age-related declines in visual abilities (acuity, contrast sensitivity, visual fields) might influence fixation selection. Previous research showed that the influence of low-level image features on fixation selection decreases with age. This study uses generalized linear mixed modeling (GLMM) to analyze age-related differences in the effects of location-based and object-based visual salience on fixation selection. The GLMM approach allows explicit modeling of central bias and thus assessing the unique contribution of salience.

Existing literature presents two contrasting views on fixation selection during scene viewing: the salience view and the object view. The salience view posits that low-level image features drive eye movements towards visually salient regions. Support for this comes from saliency map algorithms that reasonably predict human fixations. However, the object view emphasizes objects as the primary unit of saccadic selection. Studies demonstrating a preferred viewing location (PVL) near the center of objects support this view. The PVL is influenced by various factors like object size and saccade direction. The relationship between saliency and object-based selection remains debated. Some studies indicate that saliency maps primarily predict object locations rather than fixations directly. Others demonstrate that object-based models, considering PVL, predict fixations as well as or better than purely salience-based models, particularly when objects are experimentally separated from salient regions. Research on age-related differences in eye movements reveals inconsistencies, with some studies finding increased oculomotor capture in older adults and others not. Overall, the influence of low-level image features on fixation selection seems to decrease with age.

Data from 42 young adults (mean age 22.1 years) and 34 older adults (mean age 72.1 years) were analyzed. Participants viewed 150 color photographs of natural scenes for 6 seconds each while their eye movements were recorded using an EyeLink 1000 system. After 20% of the trials, participants answered yes/no questions about objects in the scenes to assess scene encoding. The Adaptive Whitening Saliency (AWS) model was used to compute salience maps. An independent annotator labeled objects in the scenes by providing bounding boxes. Two generalized linear mixed models (GLMMs) were used to analyze fixation data: a grid GLMM (analyzing fixation guidance without recurrence to objects) and an object GLMM (analyzing object prioritization for gaze guidance). The grid GLMM used a 8x6 grid to partition each image into equal-sized cells. Both GLMMs included fixed effects for central bias, object size (for object GLMM), object salience, and age group. Random effects included subject, object (for object GLMM), and scene. Additionally, two linear mixed models analyzed first-fixation duration and first-pass gaze duration. Finally, the distribution of within-object fixation locations was analyzed to assess the preferred viewing location (PVL). Normalized x- and y-coordinates were used to represent landing positions within objects relative to the object's center. Two linear mixed models were used to analyze the horizontal and vertical components of these normalized landing positions.

The analysis revealed no significant differences between age groups in terms of the number of fixations, fixation durations, or saccade amplitudes. However, GLMM analyses uncovered systematic differences. The grid GLMM showed that location-based visual salience predicted fixation probability in young adults, but this effect was significantly smaller in older adults. The object GLMM revealed that object salience significantly predicted fixation probability in both young and older adults, with a larger effect size for older adults. Object size also influenced fixation probability, with a stronger effect for older adults. Both age groups exhibited a preferred viewing location (PVL) close to the object center, with a slight tendency to undershoot. Analyses of fixation times showed that older adults had longer gaze durations on selected objects compared to younger adults. This was not reflected in first fixation durations, indicating that older adults made more object refixations. In summary, visual salience influenced fixation selection but object-based salience was more prominent in older adults.

The findings support the notion that objects are crucial units of saccadic selection in scene perception. The results reconcile the salience view with the object view, showing that while objects dominate the selection process, visual salience plays a role in prioritizing among objects. The age-related differences suggest that older adults rely more on object-bound information. The increased effect of object-based salience in older adults, despite a reduced effect of location-based salience, indicates that they might compensate for age-related declines in low-level visual processing by emphasizing higher-level object-based information. The consistent PVL across age groups suggests that extrafoveal processing remains relatively preserved. Further research should investigate the role of scene semantics and the potential interaction between top-down and bottom-up processes in scene perception across the lifespan.

This study demonstrates that object-based visual salience significantly contributes to fixation selection during natural scene viewing, especially in older adults. Visual salience plays a role in object prioritization. The findings highlight the importance of objects as units of attentional selection and reconcile contrasting views in the literature. Future research could explore the influence of task demands and semantic context on these findings, as well as investigate individual differences in gaze behavior more extensively.

One limitation is that visual abilities were not independently assessed, only relying on self-report. Another limitation is that the memory task might have influenced fixation patterns, biasing participants towards individual objects. The study also does not explicitly address scene semantics, though future work should account for this factor. Finally, the choice of AWS model might introduce some inherent bias due to its implicit inclusion of objectness information.