Supervised and unsupervised MRF based 3D scene classification in multiple view airborne oblique images
Keywords: Classification, Graph Cut, Learning, Performance, Point Cloud, Optimization, Random Trees
Abstract. In this paper we develop and compare two methods for scene classification in 3D object space, that is, not single image pixels get classified, but voxels which carry geometric, textural and color information collected from the airborne oblique images and derived products like point clouds from dense image matching. One method is supervised, i.e. relies on training data provided by an operator. We use Random Trees for the actual training and prediction tasks. The second method is unsupervised, thus does not ask for any user interaction. We formulate this classification task as a Markov-Random-Field problem and employ graph cuts for the actual optimization procedure.
Two test areas are used to test and evaluate both techniques. In the Haiti dataset we are confronted with largely destroyed built-up areas since the images were taken after the earthquake in January 2010, while in the second case we use images taken over Enschede, a typical Central European city. For the Haiti case it is difficult to provide clear class definitions, and this is also reflected in the overall classification accuracy; it is 73% for the supervised and only 59% for the unsupervised method. If classes are defined more unambiguously like in the Enschede area, results are much better (85% vs. 78%). In conclusion the results are acceptable, also taking into account that the point cloud used for geometric features is not of good quality and no infrared channel is available to support vegetation classification.