Improving superpixel boundaries using information beyond the visual spectrum

Keith Sullivan, Wallace E. Lawson, Donald Sofge

Abstract: Superpixels enable a scene to be analyzed on a larger scale, by examining regions that have a high level of similarity. These regions can change depending on how similarity is measured. Color is a simple and effective measure, but it is adversely affected in environments where the boundary between objects and the surrounding environment are difficult to detect due to similar colors and/or shadows. We extend a common superpixel algorithm (SLIC) to include near-infrared intensity information and measured distance information to help oversegmentation in complex environments. We demonstrate the efficacy of our approach on two problems: object segmentation and scene segmentation.