Matching Distance and Geometric Distribution Aided Learning Multiview Point Cloud Registration

TL;DR

This work tackles multiview point cloud registration by dividing the problem into reliable pose-graph construction and motion synchronization. It introduces a neural module that predicts pairwise overlap from descriptor-distance statistics and selects high-overlap pairs, followed by a data-driven motion synchronization network that alternates absolute and relative feature updates with a confidence-aware attention mechanism and geometric distribution cues. The approach achieves state-of-the-art or competitive results on indoor/outdoor datasets, demonstrates strong generalization when trained only on 3DMatch, and offers improved efficiency compared with traditional IRLS-based synchronization. Overall, the method provides a practical, scalable framework for robust multiview registration in diverse environments, with open-source code available at the provided GitHub link.

Abstract

Multiview point cloud registration plays a crucial role in robotics, automation, and computer vision fields. This paper concentrates on pose graph construction and motion synchronization within multiview registration. Previous methods for pose graph construction often pruned fully connected graphs or constructed sparse graph using global feature aggregated from local descriptors, which may not consistently yield reliable results. To identify dependable pairs for pose graph construction, we design a network model that extracts information from the matching distance between point cloud pairs. For motion synchronization, we propose another neural network model to calculate the absolute pose in a data-driven manner, rather than optimizing inaccurate handcrafted loss functions. Our model takes into account geometric distribution information and employs a modified attention mechanism to facilitate flexible and reliable feature interaction. Experimental results on diverse indoor and outdoor datasets confirm the effectiveness and generalizability of our approach. The source code is available at https://github.com/Shi-Qi-Li/MDGD.

Figures (5)

• Figure 1: The number of matching and the empirical cumulative distribution function (eCDF) show a proportional relationship with overlap. Conversely, the average matching distance decreases as overlap increases. The standard deviation of the matching distance follows an approximate bell curve distribution.
• Figure 2: A failure registration case. Left: Point clouds registered by the predicted transformation. Darker colors represent keypoints whose associated correspondence confirms the predicted transformation. Right: Red points indicate keypoints lying in a fitting plane, while black points represent those do not lie in the plane. Better viewing with color and zooming in.
• Figure 3: Pipeline for multiview point cloud registration with two parts: pose graph construction and motion synchronization.
• Figure 4: Motion synchronization module architecture. The pose graph is first initialized and projected to feature space. Then the alternate rotation and translation feature update is applied, both including confidence attention-based absolute update and relative update. The tilde notation indicates features after the confidence attention module, while the hat notation indicates features after the update.
• Figure 5: Qualitative comparison results. Top: 3DMatch studyroom2. Bottom: ScanNet scene0334_02.