Garstka, JensPeters, GabrieleMathias Paulin and Carsten Dachsbacher2014-12-162014-12-1620141017-4656https://doi.org/10.2312/egp.20141065We propose a reinforcement learning approach for an adaptive selection and application of 3D point cloud feature descriptors for the purpose of 3D object classification. The result of the learning process is an autonomously learned strategy of selection of descriptors with the property that the successive application of these descriptors to a 3D point cloud yields high classification rates among a large number of object classes. The order in which the descriptors are applied to an unfamiliar point cloud depends on the features calculated in previous steps of the descriptor sequence, i.e., the sequence of descriptors depends on the object to be classified, thus it is highly adaptive. Our approach starts with a given number of descriptors and object classes, but it is able to adapt dynamically to changes in the environment. For example, further descriptors can be added during the learning process, and new object classes are created autonomously if necessary.I.4.8 [Image Processing and Computer Vision]Scene AnalysisObject recognitionI.2.6 [Artificial Intelligence]LearningReinforcement LearningI.2.10 [Artificial Intelligence]Vision and Scene Understanding3D/stereo scene analysis