Pose Alignment for 3D Models and Single View Stereo Point Clouds Based on Stable Planes

This paper presents a novel method for alignment of geometrically similar 3D models. It is based on the prior that both models have at least one common tangent plane on which both can stand stably and when standing on it the models are partially aligned. To determine the final rotation around the stable plane's normal, needed for a complete alignment, we adapt an image alignment technique based on the log-polar transformation. Because the set of stable planes of a model is small enough, alignment is efficiently approached as a global optimization problem that finds the common stable plane providing the best alignment according to a similarity measure. As the method does not rely on any kind of global symmetry features, we show it can be used to register incomplete stereo point clouds of objects located on a stable plane (table, ground, etc.) with the corresponding similar 3D models. We evaluate the 3D-alignment method by comparing it to the well-known CPCA and show a significant improvement when aligning 120 models belonging to 12 different classes.