Aligning Point Clouds Using Cluster Structure

Summary & Methodology Analysis

The researchers address the instability found in traditional point matching by treating point cloud alignment and clustering as a joint optimization problem. Standard approaches often suffer when point clouds are processed independently, leading to inconsistent partitions. The authors propose Laplacian Optimal Transport (LapOT), which extends traditional entropic regularized optimal transport by adding quadratic Laplacian regularization terms. This modification forces the system to consider the cluster structure of the data when calculating how points in one set map to another. By leveraging the Sinkhorn algorithm for solving the transport problem, the method creates a robust coupling that reflects the intrinsic geometry of the datasets.

To improve consistency across point sets, the framework employs Refined Simultaneous Clustering (RSC). This process uses the couplings derived from LapOT to generate switch matrices, which are then used to update the similarity matrices for each point cloud. Once these matrices are refined, spectral clustering (a grouping method using the eigenvalues of a graph matrix) is applied to the graph Laplacians. This ensures that the resulting partitions remain consistent across different point clouds, even when the input data is affected by significant noise.

Despite these improvements, the method has practical trade-offs. The optimization of the general LapOT objective can be computationally expensive if general-purpose solvers are used without specialized algorithms, meaning developers should plan for potential performance overhead. Additionally, the paper notes that the method does not guarantee perfectly consistent clusters and that performance is dependent on the selection of hyperparameters. The research validated these techniques using the CAPOD dataset and S P 500 data, comparing outcomes against standard baseline methods.