Segmentation as a bottleneck in numerical cognition

Despite frequent appeals to the ecological relevance of number perception, number perception is typically studied using stimuli that differ greatly from real-world scenes. Consequently, it remains unclear whether numerical discrimination observed in controlled experimental settings operates under the same principles as it does 'in the wild.' In the present study, adult participants completed a number discrimination task with four stimulus conditions: (1) dot arrays, (2) dot arrays with visual noise, (3) naturalistic images, and (4) pseudo-naturalistic images designed to preserve object boundaries while minimizing featural complexity. Consistent with previous work, we found that accuracy for numerical discrimination of naturalistic and pseudo-naturalistic arrays was significantly worse than for dot arrays. Moreover, for the naturalistic conditions, but much less so for dot arrays, accuracy declined as the total number of objects increased. Crucially, this effect was driven by the presence of occluded objects, suggesting that numerical discrimination is constrained by the ability to segment individual items in cluttered visual scenes. Furthermore, in naturalistic and pseudo-naturalistic conditions, there was a significant interaction between total number of objects and object occlusion, such that the number of objects only impacted performance when stimuli contained occluded objects. Our findings extend previous work on connectedness and grouping effects in numerical perception by demonstrating that in naturalistic contexts, the demands of object segmentation, particularly when occlusion is present, create a measurable bottleneck that systematically affects numerical discrimination in ways not observed with traditional stimuli. This suggests that models of numerical cognition must account for the computational costs of resolving complex visual scenes into countable units, costs that increase with set size and visual complexity.