LEARNING TO MERGE: LEARNING IMPROVED SEGMENTATIONS FOR URBAN LIDAR DATA

Abstract

We explore methods for improving per-class segmentation of 3D point clouds of urban environments with the goal to improve semantic segmentation results. Speci cally, we implement a system which starts with an oversegmentation of a point cloud and produces per-point and per-segment semantic class labels. To aid segmentwise predictions, we present a supervised method which learns to hierarchically group segments from the oversegmentation into supersegments which can be more accurately classi ed than the individual segments alone. We evaluate this method based on improvements to the initial oversegmentation and the classi cation accuracy over alternative approaches, including a region pro- posal algorithm. We nd that it improves segmentation results, but does not always improve semantic segmentation performance.