Phutanik Setasartit¹ (), Payachana Victoria Chareunsouk^1,2, Waragon Phusuwan^1,2, Chaipat Chunharas^1,3,4; ¹Cognitive Clinical and Computational Neuroscience Center of Excellence, Faculty of Medicine, Chulalongkorn University, Bangkok, Thailand, ²Medical Sciences, Faculty of Medicine, Chulalongkorn University, Bangkok, Thailand, ³Division of Neurology, Department of Medicine, Faculty of Medicine, King Chulalongkorn Memorial Hospital, Thai Red Cross Society, Bangkok, Thailand, ⁴Chula Neuroscience Center, King Chulalongkorn Memorial Hospital, Thai Red Cross Society, Bangkok, Thailand

Everyday scenes contain multiple objects, yet we effortlessly classify them. Prior research suggests that the properties of objects in a scene, known as diagnosticity, play a crucial role in scene categorization. A single diagnostic object can be sufficient to support scene classification. However, when multiple diagnostic objects provide converging or conflicting diagnosticities, the way they are jointly interpreted remains poorly understood. We hypothesize that room classification is primarily determined by how individual object diagnosticities are co-evaluated. Specifically, we focused on four integration models: (1) Winner-take-all, relying on the most diagnostic object; (2) Interactive, in which the most unlikely object constrains the overall diagnosticities; (3) Additive, summing object diagnosticities; and (4) Average, which proportionates to the total diagnosticity. We evaluated these models against responses from nine participants who completed a two-alternative forced-choice task across 4,860 trials. In each trial, they viewed a brief visual scene containing either one or two objects, selected from nine objects across three room categories: bedroom, bathroom, and kitchen. After 300 ms, the text probe of the room type was displayed, and participants judged whether the composition matched the probe. For two-object trials, the average model provided the best fit to participant responses (r = .85), outperforming the other models (r ≈ .70–.80). To further examine the remaining model–data mismatches, we developed an extended model in which object diagnosticities were learned from the data. This substantially improved model performance (r = .93), with the largest increase observed in the bedroom condition (r = .98). Adding an interaction term resulted in a slight to no benefit (r = .94) with a lower Bayesian Information Criterion (BIC) than earlier models. These findings suggest that the weighted-sum model is most compatible with all hypotheses, with some rooms (e.g., bedroom) depending on highly diagnostic individual objects (e.g., bed).