MCI-Net: A Robust Multi-Domain Context Integration Network for Point Cloud Registration

TL;DR

MCI-Net tackles the robustness of point cloud registration by moving beyond Euclidean neighborhoods to multi-domain context integration. It introduces GNAM for global structural reasoning, PCIM for intra-domain decoupling and inter-domain fusion, and DISM for history-aware inlier weighting across iterative pose updates. The approach yields state-of-the-art performance on indoor benchmarks (e.g., 3DMatch/3DLoMatch) and strong results outdoors (KITTI Odometry), with ablations confirming the contribution of each module. This work advances robust feature learning and correspondence weighting in challenging registration scenarios, offering practical gains for 3D vision tasks.

Abstract

Robust and discriminative feature learning is critical for high-quality point cloud registration. However, existing deep learning-based methods typically rely on Euclidean neighborhood-based strategies for feature extraction, which struggle to effectively capture the implicit semantics and structural consistency in point clouds. To address these issues, we propose a multi-domain context integration network (MCI-Net) that improves feature representation and registration performance by aggregating contextual cues from diverse domains. Specifically, we propose a graph neighborhood aggregation module, which constructs a global graph to capture the overall structural relationships within point clouds. We then propose a progressive context interaction module to enhance feature discriminability by performing intra-domain feature decoupling and inter-domain context interaction. Finally, we design a dynamic inlier selection method that optimizes inlier weights using residual information from multiple iterations of pose estimation, thereby improving the accuracy and robustness of registration. Extensive experiments on indoor RGB-D and outdoor LiDAR datasets show that the proposed MCI-Net significantly outperforms existing state-of-the-art methods, achieving the highest registration recall of 96.4\% on 3DMatch. Source code is available at http://www.linshuyuan.com.

Figures (5)

• Figure 1: Registration recall (RR) on 3DLoMatch and 3DMatch with 2,500 points. The proposed method significantly outperforms all the state-of-the-art approaches.
• Figure 2: Architecture of the proposed MCI-Net. Given point clouds $P$ and $Q$, patch-level and point-level features are first extracted using the PAConv-based feature pyramid network yao2024pare. The extracted patch features are embedded into coordinate and feature spaces, and then processed by the multi-domain context integration module, where GNAM captures global structural relationships, and PCIM decouples and fuses contextual information across domains to enhance feature discriminability. The patch-to-point matching strategy is used to obtain point correspondences. Finally, the dynamic inlier selection method iteratively updates correspondence weights to progressively refine the inlier set, yielding the final pose estimation $(R, t)$.
• Figure 3: The proposed intra-domain feature decoupling. Features are decomposed into global-aligned and residual parts, and the residual is then concatenated with the global feature to capture both local details and global context.
• Figure 4: The proposed inter-domain context interaction. It contains three parallel inter-domain cross-attention (ICA) branches for contextual information fusion across domains.
• Figure 5: Qualitative results on 3DMatch and 3DLoMatch datasets. Columns (b) and (c) show the correspondences, while columns (d) and (e) demonstrate the registration results. Green/red lines indicate inliers/outliers.