Accurate Point Cloud Registration with Robust Optimal Transport

Area of Science:

• Computer Vision
• Computational Geometry
• Medical Imaging

Background:

• Shape matching is crucial for various applications.
• Existing methods for point cloud registration face challenges in accuracy and computational cost.
• Optimal Transport (OT) theory offers a powerful framework for comparing distributions, applicable to shape matching.

Purpose of the Study:

• To investigate the efficacy of robust Optimal Transport (OT) for enhancing shape matching and point cloud registration.
• To address practical challenges in applying OT to shape matching problems.
• To demonstrate state-of-the-art performance of OT-enhanced registration models.

Main Methods:

• Utilized recent OT solvers to improve optimization-based and deep learning registration methods.
• Developed solutions for difficulties in applying OT to shape matching.
• Evaluated transport-enhanced registration on rigid registration, scene flow estimation (Kitti dataset), and lung vascular tree registration.

Main Results:

• OT-based methods achieved state-of-the-art accuracy and scalability on Kitti and lung registration tasks.
• Demonstrated improved accuracy and affordable computational cost for point cloud registration.
• Introduced PVT1010, a new dataset for challenging lung vascular tree registration.

Conclusions:

• Robust OT is a key method for computer vision, enabling fast pre-alignment and fine-tuning for registration models.
• OT significantly boosts accuracy in point cloud registration at a reasonable computational expense.
• The developed OT-based approach provides a versatile and effective solution for complex shape matching problems.