Abstract
Scenes contain a wealth of information that can be selected and processed differently depending on the requirements of the observer. While some scenes are categorized by their global spatial layout (e.g. corridor) and therefore defined by their boundary-related parts, others are categorized by local diagnostic objects (e.g. bed in a bedroom) and hence are defined by their content-related parts. Additionally, the spatial scale of information required to identify and categorize visual stimuli has been shown to change (Collin & McMullen, 2005, Percept Psychophys 6: 354) depending on the level of classification (subordinate, basic, superordinate). Several studies that have attempted to characterize scenes based on hierarchy of components and spatial scale often cue the participant to respond to one of these specific levels, and therefore bias the perception of the scene overall. We conducted an extensive investigation using an uncued, and therefore unbiased, task in which participants freely described scenes at multiple spatial scales. Independent observers subsequently rated the responses for accuracy on a predetermined list of attributes shown to tap into hierarchical categorization (similar to Fei-Fei et al., 2007, J Vis 7:1). Data were analyzed to identify the scale of interest, which is described as the largest transition in accuracy across the spatial scales. Preliminary results reveal that the scale of interest follows a hierarchical progression, such that superordinate categories of scene and object are more accurately described at lower spatial scales than basic or subordinate categories. However, no significant difference was observed between the later two. In addition, participant's perceived responses were compared to the ground-truth for several attribute categories revealing a bias to report low frequency images as outdoor. Results are discussed with reference to perceptual differences in natural/urban environments, sensory and semantic perception, and the taxonomy of scene categorization.
Meeting abstract presented at VSS 2014