Point sets joint registration and co-segmentation

Abstract

We present a novel approach of joint registration and co-segmentation for point sets where objects move in different ways. We consider joint registration and co-segmentation as two problems that are heavily entangled with each other; thus, we represent the input point sets as samples from a generative model and bring up with a novel formulation based on Gaussian mixture model. By maximizing the posterior probability of the samples, we gradually recover the latent object models as well as an object-level segmentation and simultaneously align the segmented points to the latent object models. Along with the formulation, we design an interactive tool that helps users intuitively intervene the process to optimize the registration and segmentation results. The experiment results on a group of synthetic and scanned point clouds demonstrate that our method is powerful and effective for joint registration and co-segmentation on point sets of multiple objects.