Abstract:This paper presents a novel non-rigid point set registration method that is inspired by unsupervised clustering analysis. Unlike previous approaches that treat the source and target point sets as separate entities, we develop a holistic framework where they are formulated as clustering centroids and clustering members, separately. We then adopt Tikhonov regularization with an $\ell_1$-induced Laplacian kernel instead of the commonly used Gaussian kernel to ensure smooth and more robust displacement fields. Our formulation delivers closed-form solutions, theoretical guarantees, independence from dimensions, and the ability to handle large deformations. Subsequently, we introduce a clustering-improved Nyström method to effectively reduce the computational complexity and storage of the Gram matrix to linear, while providing a rigorous bound for the low-rank approximation. Our method achieves high accuracy results across various scenarios and surpasses competitors by a significant margin, particularly on shapes with substantial deformations. Additionally, we demonstrate the versatility of our method in challenging tasks such as shape transfer and medical registration.

What problem does this paper attempt to address?

The problem that this paper attempts to solve is the challenges in non - rigid point set registration, especially when dealing with shapes with large deformations. Non - rigid point set registration refers to finding a non - linear displacement field to accurately align one geometric shape with another. This problem is very important in multiple applications such as scene reconstruction, pose tracking, animation, deformable shape manipulation and editing. However, due to the increased degrees of freedom of allowed deformations, especially when dealing with shapes with significant deformations, non - rigid registration is more challenging than rigid registration. The main contribution of the paper lies in proposing a new, correspondence - free method to solve the non - rigid point set registration problem, with special emphasis on dealing with point sets with significant deformations. This method achieves this goal by reformulating the non - rigid deformation process as an unsupervised clustering problem in machine learning. Specifically, this method regards the source point set as the cluster centers and the target point set as the cluster samples, so that it can utilize the interaction between the two. In this way, the process of dynamically optimizing and updating the cluster centers corresponds to the underlying deformation of the source shape. In addition, this method also introduces the sparsity - inducing Laplacian kernel (ℓ1 - norm) for use in Tikhonov regularization to ensure that the displacement field of the cluster centers is as smooth as possible, and provides a strict theoretical analysis to prove the approximation error bound of the Nyström low - rank method. In summary, this paper aims to provide a novel, correspondence - free non - rigid point set registration method that performs well when dealing with shapes with significant deformations and verifies its superiority in multiple experiments.

Correspondence-Free Non-Rigid Point Set Registration Using Unsupervised Clustering Analysis

Non-rigid Point Set Registration Based on Gaussian Mixture Model with Integrated Feature Divergence

Graph Correspondence-Based Point Set Registration

Robust Non-Rigid Point Set Registration Using Student's-T Mixture Model.

Clique-like Point Cloud Registration: A Flexible Sampling Registration Method Based on Clique-like for Low-Overlapping Point Cloud

Non-Rigid Point Set Registration with Robust Transformation Estimation under Manifold Regularization

Non-Rigid 3D Point Set Registration With Reliable Hybrid Mixture Model.

Non-rigid point set registration based on local neighborhood information support

Accurate Registration of Cross-Modality Geometry via Consistent Clustering

Bidirectional Patch-based Correlations with Local Rigidity for Global Nonrigid Registration

Non-Rigid Point Set Registration Using Multi-Feature and Gaussian Mixture Model

Nonrigid Point Set Registration by Preserving Local Connectivity.

Learning Sufficient Correlations among Points for 3D Non-rigid Point Cloud Registration

Nonrigid Point Set Registration With Robust Transformation Learning Under Manifold Regularization.

An Efficient Volumetric Method for Non-Rigid Registration

Point Set Registration Via Rigid Transformation Consensus

SCM: Spatially Coherent Matching With Gaussian Field Learning for Nonrigid Point Set Registration

AMCNet: Adaptive Matching Constraint for Unsupervised Point Cloud Registration.

Dependent landmark drift: robust point set registration with a Gaussian mixture model and a statistical shape model

Multi-instance Point Cloud Registration by Efficient Correspondence Clustering

Efficient and Robust Point Cloud Registration via Heuristics-guided Parameter Search