Clustering of Affective Dimensions in Pictures: An exploratory analysis of the NAPS database

Photographs and other multimedia documents can easily provoke strong emotional reactions. While knowledge discovery based on semantics and visual properties of pictures is considerably explored, little attention is given to mining of affective features of visual stimuli. We report result from clustering of 1356 stimuli from the Nencki Affective Picture System database using two- dimensional (valence/arousal) model of affect and k-means unsupervised learning algorithm. Clustering with k=1–94 centroids is reported, together with dominant picture keywords and descriptive statistical parameters. Optimal number of centroids was estimated using minimum cumulative error rule. A custom Java application integrated with WEKA machine learning software was developed for the study. The results are freely available by contacting the first author. A practical goal of this research is to develop an intelligent expert system that could infer emotion content of multimedia from evaluation of its semantics, and vice versa, estimate dominant semantics from affect if such information is available. The system could potentially have many useful applications such as supported construction of affective multimedia databases, video recommendation or emotion estimation. Reported information on clustering of emotions is essential for success of such a system.