Moment-Preserving Techniques for 3D Shape Registration and Recognition

Abstract

In applying the moment-preserving technique to the problem of 3D shape recognition and registration, we develop two new and powerful improvements to this popular inference method. The first, 3D shape registration, solves the problem of 3D point-to-point correspondences with weighting estimates by preserving the 3D moments between two input shapes. The second, dynamic 3D shape modelling, explains away 3D points of high variations by shape registration, substantially reducing the number of modes in the posterior distribution of class and pose, and improving precision. Our experiments demonstrate that these contributions make the 3D shape models not only tractable but also highly accurate for our example application. As compared with the standard Hough transform, the problem of exponential memory requirements in 3D object detection, segmentation and recognition has been solved.