Clustering-Based Latent Variable Models for Monocular Non-rigid 3D Shape Recovery

Quan Wang, Fei Wang, Daming Li, Xuan Wang

Abstract: The difficulty of monocular non-rigid 3D reconstruction using statistical learning approaches is to get a model that can represent as many deformations as possible. Given a known dataset to learn a model, existing latent variable models (LVMs) fail to focus on how to attain labeled samples. In this paper, we propose novel clustering-based LVMs in which we automatically select representative samples to be the labeled ones. To this end, G-means algorithm is adopted to cluster latent variables and obtain the labeled samples. These labeled samples are corresponding to the latent variables closest to clustering centers. We learn the Gaussian Process Latent Variable Model (GPLVM) and the Constrained Latent Variable Model (CLVM) into which we introduce clustering in the context of monocular non-rigid 3D reconstruction, and compare them to those without clustering. The experimental results show that our clustering-based LVMs could perform better.