Self-Learning 3D Object Classification

We present a self-learning approach to object classification from 3D point clouds. Existing 3D feature descriptors have been utilized successfully for 3D point cloud classification. But there is not a single best descriptor for any situation. We extend a well-tried 3D object classification pipeline based on local 3D feature descriptors by a reinforcement learning approach that learns strategies to select point cloud descriptors depending on qualities of the point cloud to be classified. The reinforcement learning framework learns autonomously a strategy to select feature descriptors from a provided set of descriptors and to apply them successively for an optimal classification result. Extensive experiments on more than 200.000 3D point clouds yielded higher classification rates with partly more reliable results than a single descriptor setting. Furthermore, our approach proved to be able to preserve classification strategies that have been learned so far while integrating additional descriptors in an ongoing classification process.