Category systems for real-world scenes.

Area of Science:

• Cognitive Science
• Computer Vision
• Psychology

Background:

• Scene recognition is crucial for understanding visual perception.
• Current scene categorization methods often rely on arbitrary labels, not reflecting human perception.
• There is a need for data-driven approaches to define objective scene categories.

Purpose of the Study:

• To introduce Clustering by Increasing the Rand Index via Coordinate Ascent (CIRCA), a novel unsupervised method for deriving ground-truth scene categories.
• To establish human-preferred scene taxonomies based on semantic content, spatial structure, and visual appearance.
• To validate the generalizability and superiority of CIRCA-derived categories.

Main Methods:

• Human participants categorized 80 outdoor scenes from the Southampton-York Natural Scenes (SYNS) dataset based on meaning, 3D structure, and 2D appearance.
• The CIRCA algorithm was used to determine optimal category structures and derive labels for each dimension.
• Experiments involved generalization tests, comparisons with the spatial envelope model, and validation on an independent dataset.

Main Results:

• CIRCA successfully derived representative scene category structures and labels from human judgments.
• The derived categories generalized well to new images and observers.
• CIRCA-based category systems outperformed the SUN taxonomy and an alternative clustering method.

Conclusions:

• CIRCA offers an objective, data-driven method for establishing scene categories from psychophysical data.
• This approach can generate more accurate and human-aligned taxonomies for scene recognition research.
• CIRCA has broad applicability for deriving optimal categorical groupings across various datasets.