Detection and Segmentation of Quasi-Planar Surfaces Through Expectation Maximization Under a Planar Homography Constraint

We propose a novel method capable of detecting and segmenting quasi-planar surfaces based on homograph decomposition and Semi-Global-Matching without the need for extrinsic calibration or stereo rectification. Existing methods require co planarity of all points on the dominant plane and are thus unsuited for unconstrained quasi-planar surfaces. In contrast to state of the art methods we can account for local depth variations. This is achieved by introducing a novel planar inter-image rectification technique, which enables us to perform block matching without popular rectification. Experiments are performed on two different databases. Firstly we quantitatively evaluate the general feasibility of our method on a database containing indoor and outdoor scenes with available ground truth data. Secondly we apply our method to the new publicly available HAWK wood database. Our experiments have shown that the true positive rate of our segmentation procedure exceeds 94.0% while the false positive rate is below 4.9%.