Rectified Point Flow: Generic Point Cloud Pose Estimation

2025-06-05

AI Summary

This work addresses two core challenges in point cloud pose estimation—pairwise registration and multi-part shape assembly—via a unified conditional generative framework that implicitly learns intrinsic symmetry structure without requiring manual symmetry annotations. The method employs neural ordinary differential equations (ODEs) to model continuous point-wise velocity fields, driving noisy points from pose-agnostic input clouds toward their target positions while simultaneously recovering the full six-degree-of-freedom poses of all parts. A self-supervised overlap point encoder is introduced to enforce local geometric consistency. Key contributions include: (i) the first implicit symmetry learning paradigm that operates without explicit symmetry labels; and (ii) a joint training scheme for registration and assembly that leverages shared geometric priors. Evaluated on six standard benchmarks, the approach achieves state-of-the-art performance, with particularly notable improvements in pose estimation accuracy for complex symmetric objects and generalization across non-rigid deformations.

Technology Category

• Intelligent Robots: State Estimation
• Computer Vision: Visual Reasoning & Symbolic Representations
• Machine Learning: Deep Generative Models & Autoencoders

Application Category

• Semantics and Knowledge: Methods to enhance, augment, integrate or synergize semantic models such as knowledge graphs and LLMs
• Graph Algorithms and Modeling for the Web: Graph embeddings and representation learning for Web-related graphs
• Search and Retrieval-Augmented AI: Web learning to rank, online learning, and counterfactual learning for ranking

Abstract

We introduce Rectified Point Flow, a unified parameterization that formulates pairwise point cloud registration and multi-part shape assembly as a single conditional generative problem. Given unposed point clouds, our method learns a continuous point-wise velocity field that transports noisy points toward their target positions, from which part poses are recovered. In contrast to prior work that regresses part-wise poses with ad-hoc symmetry handling, our method intrinsically learns assembly symmetries without symmetry labels. Together with a self-supervised encoder focused on overlapping points, our method achieves a new state-of-the-art performance on six benchmarks spanning pairwise registration and shape assembly. Notably, our unified formulation enables effective joint training on diverse datasets, facilitating the learning of shared geometric priors and consequently boosting accuracy. Project page: https://rectified-pointflow.github.io/.

Problem

• Unified parameterization for point cloud registration and shape assembly
• Learning continuous velocity field for point transport without symmetry labels
• Achieving state-of-the-art performance on multiple benchmarks via joint training

Innovation

• Learns continuous point-wise velocity field
• Intrinsically learns assembly symmetries
• Self-supervised encoder for overlapping points