Unsupervised 3D Object Discovery and Categorization for Mobile Robots

We present a method for mobile robots to learn the concept of objects and categorize them without supervision using 3D point clouds from a laser scanner as input. In particular, we address the challenges of categorizing objects discovered in different scans without knowing the number of categories. The underlying object discovery algorithm finds objects per scan and gives them locally-consistent labels. To associate these object labels across all scans, we introduce class graph which encodes the relationship among local object class labels. Our algorithm finds the mapping from local class labels to global category labels by inferring on this graph and uses this mapping to assign the final category label to the discovered objects. We demonstrate on real data our alogrithm’s ability to discover and categorize objects without supervision.