Image databases are central to empirical aesthetics, enabling tests of how image statistics relate to observers’ appreciation. However, many existing databases have two key limitations: (1) they conflate low-level visual features with high-level semantic content, making it difficult to separate visual from cognitive influences on aesthetic judgments; and (2) they are imbalanced, overrepresenting highly appreciated images. To address these issues, we present the Minimum Semantic Content (MSC) database, a large, systematically curated resource for computational aesthetics. It comprises 10,426 natural scenes with reduced, homogenized semantic content, minimizing cognitive and emotional confounds. Each received 100 individual aesthetic ratings from naïve observers, drawn from a pool of approximately 10,000 participants, via crowdsourcing. The database includes both “beautified” and “uglified” versions, generated with a manipulation technique that promotes uniform coverage across the aesthetic spectrum. This broader distribution mitigates bias and overfitting in models. Validation also shows improved robustness in computational models overall. This database enables researchers to study how perceptual features shape aesthetic judgments, using stimuli with very limited semantic and contextual confounds.
Penacchio et al. (Tue,) studied this question.