Graph Matching Optimization Network for Point Cloud Registration

TL;DR

GMONet tackles partial point cloud registration by enforcing isometry-preserving constraints during feature learning through two graph-matching optimizers. It integrates a coarse-level Partial Graph Matching Optimizer on down-sampled super points with a fine-level Mini-batch Full Graph Matching Optimizer on the overlap region, accelerated by the inexact proximal point method and mini-batch optimal transport, and uses a KPConv-based backbone with geometric attention to produce discriminative, geometry-consistent features. The training losses combine a coarse overlap-aware circle loss, a coarse-level PMMO loss, a fine-level mini-batch graph matching loss, and an overlap supervision term, guided by balanced weighting. Evaluations on the indoor 3DMatch/3DLoMatch and outdoor KITTI benchmarks show GMONet achieving competitive registration recall and improved geometric accuracy, validating that explicit isometry constraints in feature learning enhance overlap detection and rigid matching. The approach offers a scalable way to incorporate rigorous geometric constraints into learning-based registration, potentially benefiting robust pose estimation in real-world, partially overlapping scenes.

Abstract

Point Cloud Registration is a fundamental and challenging problem in 3D computer vision. Recent works often utilize the geometric structure information in point feature embedding or outlier rejection for registration while neglecting to consider explicitly isometry-preserving constraint ($e.g.,$ point pair linked edge's length preserving after transformation) in training. We claim that the explicit isometry-preserving constraint is also important for improving feature representation abilities in the feature training stage. To this end, we propose a \underline{G}raph \underline{M}atching \underline{O}ptimization based \underline{Net}work (GMONet for short), which utilizes the graph-matching optimizer to explicitly exert the isometry preserving constraints in the point feature training to improve the point feature representation. Specifically, we exploit a partial graph-matching optimizer to optimize the super point ($i.e.,$ down-sampled key points) features and a full graph-matching optimizer to optimize fine-level point features in the overlap region. Meanwhile, we leverage the inexact proximal point method and the mini-batch sampling technique to accelerate these two graph-matching optimizers. Given high discriminative point features in the evaluation stage, we utilize the RANSAC approach to estimate the transformation between the scanned pairs. The proposed method has been evaluated on the 3DMatch/3DLoMatch benchmarks and the KITTI benchmark. The experimental results show that our method performs competitively compared to state-of-the-art baselines.

Figures (2)

• Figure 1: Overview of our proposed GMONet. First, the point clouds $\mathbf{P}$ and $\mathbf{Q}$ are fed to the down-sampling encoder and geometric attention layers to obtain the super points ($\hat{\mathbf{P}}$,$\hat{\mathbf{Q}}$), their features ($\mathbf{F}^{\hat{P}}$,$\mathbf{F}^{\hat{Q}}$), and overlapping scores ($\mathbf{O}^{\hat{P}}$, $\mathbf{O}^{\hat{Q}}$). Then, we apply the partial graph matching optimizations on super points to improve the overlap region detecting ability. Next, we use three upsampling layers to recover the fine-level points, their features ($\mathbf{F}^{P}$,$\mathbf{F}^{Q}$), and overlapping scores ($\mathbf{O}^{P}$, $\mathbf{O}^{Q}$). Lastly, a mini-batch graph matching optimizer is applied on the fine-level points in the overlap region to enhance the features' abilities for global "rigid" matching.
• Figure 2: Visualization of the effective role of coarse-level partial graph matching constraint and fine-level graph matching constraint.

Theorems & Definitions (1)

• Definition 1