Abstract

3D registration, a crucial process to establish correspondences between surfaces, often relies on traditional mesh alignment methods that primarily employ ICP-based (Iterative Closest Point) optimization for non-rigid deformation. While the embedded deformation method was designed to expedite this process, enabling a plethora of real-time applications, it does pose challenges. Its reliance on a simplified structure for regularization can be problematic when dealing with intricate cases where this structure fails to fully encapsulate the surface attributes. Furthermore, without meticulous parameter tuning, deformation often underperforms due to slow convergence or being trapped in a local minimum, especially when all regions on the surface are assumed to have uniform rigidity during optimization.

In this research, we put forth a groundbreaking solution that uncouples regularization from the underlying deformation model, by effectively managing the rigidity of vertex clusters. We devised a swift, two-step solution that alternates between isometric deformation and embedded deformation, incorporating cluster-based regularization. Our method readily supports region-adaptive regularization with cluster refinement, delivering efficient execution. Comprehensive experiments validate the efficacy of our approach for mesh alignment tasks, even under challenging conditions of large-scale deformation and imperfect data. Demonstrably superior both numerically and visually.

IEEE Xplore

Recommended citation: K. Chen, F. Yin, B. Du, B. Wu and T. Q. Nguyen, "Efficient Registration for Human Surfaces Via Isometric Regularization on Embedded Deformation," in IEEE Transactions on Visualization and Computer Graphics, 2022