UM3: Unsupervised Map to Map Matching

TL;DR

UM3 tackles map-to-map matching without labeled data by introducing pseudo coordinates that capture the relative spatial layout of nodes, and by unifying feature-based and geometry-based similarities within an unsupervised optimization framework. A geometric-consistent loss together with a tile-based extension enables scalable, boundary-coherent matching on large-scale maps. The approach employs a GNN backbone to learn node embeddings and uses Sinkhorn normalization with a Hungarian post-processing step to produce hard correspondences, achieving state-of-the-art accuracy and robust performance under noise. This work offers a practical, scalable solution for integrating heterogeneous geospatial datasets and has significant implications for map alignment, navigation, and urban analytics.

Abstract

Map-to-map matching is a critical task for aligning spatial data across heterogeneous sources, yet it remains challenging due to the lack of ground truth correspondences, sparse node features, and scalability demands. In this paper, we propose an unsupervised graph-based framework that addresses these challenges through three key innovations. First, our method is an unsupervised learning approach that requires no training data, which is crucial for large-scale map data where obtaining labeled training samples is challenging. Second, we introduce pseudo coordinates that capture the relative spatial layout of nodes within each map, which enhances feature discriminability and enables scale-invariant learning. Third, we design an mechanism to adaptively balance feature and geometric similarity, as well as a geometric-consistent loss function, ensuring robustness to noisy or incomplete coordinate data. At the implementation level, to handle large-scale maps, we develop a tile-based post-processing pipeline with overlapping regions and majority voting, which enables parallel processing while preserving boundary coherence. Experiments on real-world datasets demonstrate that our method achieves state-of-the-art accuracy in matching tasks, surpassing existing methods by a large margin, particularly in high-noise and large-scale scenarios. Our framework provides a scalable and practical solution for map alignment, offering a robust and efficient alternative to traditional approaches.

Figures (5)

• Figure 1: Overview of Unsupervised Map-to-Map Matching. The key steps include: (1) transforming raw GPS coordinates into pseudo coordinates that preserve spatial topology; (2) learning a correspondence matrix through the fusion of feature and geometric similarity; and (3) optimizing a map-to-map matching–specific loss that incorporates spatial and structural constraints. These steps correspond to Sections \ref{['sec: pseudo coordinates construction']}, \ref{['sec: learning node corespondence']}, and \ref{['sec: unsupervised loss function']}, respectively.
• Figure 2: Visualization results of our method on real world datasets. Since the source and target maps are from the same region, the corresponding roads generally maintain similar relative positions across maps. The correctly matched roads are highlighted, while roads shown in thin black lines indicate unmatched or mismatched segments. Here, unmatched refers to roads that have no corresponding counterparts in the other map (e.g., they do not exist in the other map).
• Figure 3: Node matching result of our method on Bremen dataset. The source and target maps were collected in 2014 and 2025, respectively. A zoomed-in view of a portion of the matching results is provided as an illustrative example.
• Figure 4: Node matching result of our method on Boston-L dataset. Each light blue line connects a pair of matched nodes from the two maps. Zoom in to better view.
• Figure 5: Parameter analysis on the Shanghai dataset. (a) and (b) show the impact of varying parameters $\alpha$ and $\lambda$, respectively. The x-axis in both plots corresponds to different parameter values, and the y-axis denotes the matching accuracy. (c) illustrates the training dynamics, showing how the loss decreases and the accuracy improves over training steps.

Theorems & Definitions (2)

• Definition 1
• Definition 2