Efficient incremental map segmentation in dense RGB-D maps

Abstract

In this paper we present a method for incrementally segmenting large RGB-D maps as they are being created. Recent advances in dense RGB-D mapping have led to maps of increasing size and density. Segmentation of these raw maps is a first step for higher-level tasks such as object detection. Current popular methods of segmentation scale linearly with the size of the map and generally include all points. Our method takes a previously segmented map and segments new data added to that map incrementally online. Segments in the existing map are re-segmented with the new data based on an iterative voting method. Our segmentation method works in maps with loops to combine partial segmentations from each traversal into a complete segmentation model. We verify our algorithm on multiple real-world datasets spanning many meters and millions of points in real-time. We compare our method against a popular batch segmentation method for accuracy and timing complexity.