Abstract

In this paper we present an approach based on scene- or part-graphs for geometrically categorizing touching and
occluded objects. We use additive RGBD feature descriptors and hashing of graph configuration parameters for describing the spatial arrangement of constituent parts. The presented experiments quantify that this method outperforms our earlier part-voting and sliding window classification. We evaluated our approach on cluttered scenes, and by using a 3D dataset containing over 15000 Kinect scans of over 100 objects which were grouped into general geometric categories. Additionally, color, geometric, and combined features were compared for categorization tasks.