Leap segmentation for recovering image surface layout

Abstract

High-level vision applications often incorporate image segmentation techniques into their preprocessing stages to reduce image data and to improve overall execution efficiency. Traditional segmentation approaches often focus on creating homogenous, connected regions of pixels to roughly correspond with image object boundaries. These methods tend to blend or remove important image details and are often computationally expensive. We describe a new, highly efficient image segmentation technique - called leap segmentation - that builds a new image representation where individual pixel data is replaced with a map of chromatic- and illumination-similar regions that are adjacent but not necessarily contiguous. We show that applying this novel view of image segmentation can significantly improve the overall performance of a high-level image labeling task. We provide a detailed comparison of the leap segmentation approach with related, existing segmentation methods. We find that leap segmentation is able to achieve high accuracy results in the task of single-image labeling for surface layout reconstruction, while exhibiting execution time improvements of 10x - 15x over existing segmentation approaches.