UMERegRobust - Universal Manifold Embedding Compatible Features for Robust Point Cloud Registration

TL;DR

This work tackles robust rigid point-cloud registration under partial overlap and heterogeneous sampling by embracing Universal Manifold Embedding (UME). It introduces UMERegRobust, a complete pipeline with a UME-compatible coloring function, a Sampling Equalizer Module (SEM), and a Dense Feature Extractor to produce transformation-invariant descriptors, combined with a Matched Manifold Detector and RTUME-based hypothesis generation. Training uses three loss components—Point-wise Contrastive Loss, UME Contrastive Loss, and a Registration Loss—and hypothesis selection via PC-FCHT, yielding a RANSAC-free, end-to-end registration approach. Empirical results on outdoor KITTI/RotKITTI and nuScenes, as well as indoor 3DMatch, show state-of-the-art performance under strict precision and exceptional robustness to large rotations, underscoring the method’s practical value for SLAM and autonomous navigation.

Abstract

In this paper, we adopt the Universal Manifold Embedding (UME) framework for the estimation of rigid transformations and extend it, so that it can accommodate scenarios involving partial overlap and differently sampled point clouds. UME is a methodology designed for mapping observations of the same object, related by rigid transformations, into a single low-dimensional linear subspace. This process yields a transformation-invariant representation of the observations, with its matrix form representation being covariant (i.e. equivariant) with the transformation. We extend the UME framework by introducing a UME-compatible feature extraction method augmented with a unique UME contrastive loss and a sampling equalizer. These components are integrated into a comprehensive and robust registration pipeline, named UMERegRobust. We propose the RotKITTI registration benchmark, specifically tailored to evaluate registration methods for scenarios involving large rotations. UMERegRobust achieves better than state-of-the-art performance on the KITTI benchmark, especially when strict precision of (1°, 10cm) is considered (with an average gain of +9%), and notably outperform SOTA methods on the RotKITTI benchmark (with +45% gain compared the most recent SOTA method).

Figures (9)

• Figure 1: Top: Registration Recall (RR) performance of different baselines on regular KITTI ($x$-axis) and RotKITTI ($y$-axis) registration benchmarks. UMERegRobust outperforms the compared SOTA methods, on both benchmarks. Bottom: Registration problem from RotKITTI benchmark, highlighting significant rotation between measurements. Source and target point clouds are shown in different colors, with arrow direction representing vehicle heading, indicating a $120^\circ$ rotation problem.
• Figure 2: UMERegRobust Overview.${\mathcal{P}}, {\mathcal{Q}}$ are the input point clouds; (1) The UME Compatible Coloring Module resamples the point clouds into a uniform grid generating $\widetilde{{\mathcal{P}}}$ and $\widetilde{{\mathcal{Q}}}$, then assigns each point with a transformation invariant feature vector creating the colored point clouds ${\mathcal{F}}_{\widetilde{{\mathcal{P}}}}, {\mathcal{F}}_{\widetilde{{\mathcal{Q}}}}$. (2) Local UME descriptors, generated on a down-sampled versions of the point clouds $\widehat{{\mathcal{P}}}, \widehat{{\mathcal{Q}}}$, are denoted by $\{{\bf H}_{{\boldsymbol p}}\}_{k=1}^K,\{{\bf H}_{{\boldsymbol q}}\}_{k=1}^K$, respectively. A Matched Manifold Detector identifies corresponding local UME descriptors, forming a set of $K$ putative matched pairs ${\mathcal{C}}$. For each matched pair, an estimated transformation is obtained using the RTUME estimator. (3) Feature correlation is used to select the hypnosis that maximize the feature correlation between the point clouds.
• Figure 2: 3DMatch Benchmark
• Figure 3: PR-Curve of Keypoint matching performance under rigid transformation, partial overlap and sampling variations of UME Local descriptor vs. other descriptors.
• Figure 4: Registration results using UMERegRobust. The red arrows depict the vehicle direction at the time the LiDAR scans were acquired. The relative rotation between point clouds ($\theta$), and registration rotation and translation errors (RRE and RTE) are given for each of the examples. Best viewed zoomed in.