Year: 2025
Type: article
Abstract: . Optimal transport seeks the most efficient way to transform one probability distribution into another, typically under constraints that preserve total mass. However, in many practical applications, such as point cloud and image co-registration, the source and target data distributions may have unequal masses. Herein, this is overcome by relaxing the aforementioned constraints. This proposed “partial” optimal transport framework adaptively selects and matches subsets of both source and target distributions, thereby enabling robust outlier rejection and noise reduction. The method relaxes the strict constraints typically used in linear programming formulations of optimal transport, thus allowing flexibility on both sides of the matching problem. The resulting transport plan is then refined through branch-and-cut and mass bounding procedures that enforce binary mass assignments and further prune undesired points. For scalability, traditional linear programming solvers are replaced by an efficient gradient-based algorithm. This approach is validated on synthetic two-dimensional data and real three-dimensional point cloud data.
Authors: Irvin Chadraa, Esteban Tabak, Debra F. Laefer
Language: English
Subfield: Applied Mathematics
Field: Mathematics
Domain: Physical Sciences
Open Access status: diamond
Funder: Office of Naval Research
OpenAlex ID: W4414317625
FWCI:
Cites:
Cited by:
Related to: