Global 3D Non-Rigid Registration of Deformable Objects Using a Single RGB-D Camera.

Area of Science:

• Computer Vision
• 3D Reconstruction
• Geometric Modeling

Background:

• Accurate 3D reconstruction of dynamic objects with non-rigid deformations is challenging.
• Existing methods often require template priors or extensive data, and struggle with pose changes or camera motion.
• Problems like loop closure and shape shrinkage can degrade reconstruction quality.

Purpose of the Study:

• To develop a novel global non-rigid registration method for dynamic 3D objects.
• To enable high-quality reconstruction despite large deformations, pose variations, and camera motion.
• To reduce data requirements and avoid common registration pitfalls.

Main Methods:

• A global alignment optimization framework is used to simultaneously compute deformations for all scans.
• An as-rigid-as-possible constraint is incorporated to prevent shape shrinkage, particularly at open boundaries.
• A coarse-to-fine multi-resolution scheme is employed to manage large-scale problems and avoid local minima.

Main Results:

• The method successfully handles large non-rigid deformations and significant pose/camera variations.
• It requires only a sparse set of scans, eliminating the need for a template prior.
• Evaluations on public and real RGB-D datasets show superior performance compared to state-of-the-art methods.

Conclusions:

• The proposed global non-rigid registration method offers a robust and efficient solution for dynamic 3D object reconstruction.
• It overcomes limitations of existing approaches by addressing large deformations, data scarcity, and common registration issues.
• The method demonstrates significant improvements in accuracy and quality for dynamic 3D scene analysis.