Tell me what do you see: Detecting perceptually-separable visual patterns via clustering of image-space features in visualizations

Visualization helps users infer structures and relationships in the data by encoding information as visual features that can be processed by the human visual-perceptual system. However, users would typically need to expend significant effort to scan and analyze a large number of views before they can begin to recognize relationships in a visualization. We propose a technique to partially automate the process of analyzing visualizations. By deriving and analyzing image-space features from visualizations, we can detect perceptually-separable patterns in the information space. We summarize these patterns with a tree-based meta-visualization and present it to the user to aid exploration. We illustrate this technique with an example scenario involving the analysis of census data.