Incremental Multiview Point Cloud Registration

TL;DR

This work tackles multiview point cloud registration for unordered scans, a problem where global approaches can be brittle and drift-prone. It introduces an incremental pipeline that first builds a sparse scan graph via scan retrieval and geometric verification, then incrementally registers scans into a canonical frame with track-based aggregation and Bundle Adjustment. A key contribution is a track refinement module that enables detector-free matchers to participate by constructing coarse Tracks and refining keypoints before final optimization. Experiments on 3D(Lo)Match, ScanNet, and ETH show strong performance in accuracy, robustness, and completeness, highlighting the method's practicality for large-scale 3D reconstruction and mapping tasks.

Abstract

In this paper, we present a novel approach for multiview point cloud registration. Different from previous researches that typically employ a global scheme for multiview registration, we propose to adopt an incremental pipeline to progressively align scans into a canonical coordinate system. Specifically, drawing inspiration from image-based 3D reconstruction, our approach first builds a sparse scan graph with scan retrieval and geometric verification. Then, we perform incremental registration via initialization, next scan selection and registration, Track create and continue, and Bundle Adjustment. Additionally, for detector-free matchers, we incorporate a Track refinement process. This process primarily constructs a coarse multiview registration and refines the model by adjusting the positions of the keypoints on the Track. Experiments demonstrate that the proposed framework outperforms existing multiview registration methods on three benchmark datasets. The code is available at https://github.com/Choyaa/IncreMVR.

Figures (12)

• Figure 1: Results of our incremental registration. The input consists of multiple unordered scans captured from different viewpoints, while the output is a single aligned result representing the reconstructed scene in a uniform coordinate system. Please zoom in to see sharp geometry details in the rightmost areas.
• Figure 2: Overview of the proposed method. Our method has three components: scan graph generation (\ref{['sec:scene_graph']}), incremental scan registration (\ref{['sec:scan_reg']}) and track refinement (\ref{['sec:track_ref']}, optionally for detector-free matchers).
• Figure 3: The process of sparse scan graph generation. The sparse scan graph is established by leveraging both scan retrieval and geometric verification.
• Figure 4: Incremental Scan Registration. Given a scan graph with relative transformations, the incremental process first registers the initial pair into the global coordinate system and creates initial tracks. Then, it iteratively registers the next scan and manages the track data through continue, create, and filtering operations. The Bundle Adjustment is performed to optimize tracks throughout the entire process.
• Figure 5: Aggregation. Two sets of point cloud frames (colored by purple and green), share a common scanning area. Corresponding keypoints from each set are merged into a landmark point in the global coordinate system through an aggregation operation.