MapData: Global Map-Image Registration

• MapData is a multimodal remote-sensing dataset defined for precise geometric registration between vector-based maps and visible imagery, featuring 121,781 aligned pairs across 233 sites.
• It employs a hybrid manual–automated pipeline, starting with manual control-point annotation and refined by iterative template matching to achieve sub-pixel homography accuracy.
• The dataset underpins robust evaluation benchmarks for cross-modal registration and facilitates the development of advanced methods such as MapGlue.

MapData is a multimodal remote-sensing dataset introduced together with the MapGlue framework for electronic map–visible image matching. It is designed for cross-modal registration between vector-based rasterized electronic navigation maps and visible-spectrum satellite or aerial imagery, with emphasis on scale, geographic diversity, and homography-based alignment quality. In the formulation reported for MapGlue, MapData spans 233 uniformly distributed geographic sampling points across all continents, provides 121,781 aligned map–visible image pairs after cleaning, and supplies a single ground-truth homography matrix $H$ for each $512\times512$ pair, thereby supporting geometric registration and localization benchmarks under controlled perturbation regimes (Wu et al., 20 Mar 2025).

1. Definition and dataset scope

MapData is a benchmark dataset for multimodal remote-sensing image matching, specifically the alignment of electronic navigation maps with visible-spectrum imagery. Its two modalities are electronic navigation maps, described as vector-based raster tiles from the Google Maps API, and visible-spectrum satellite or aerial images, described as raster imagery from Google Earth Engine (Wu et al., 20 Mar 2025).

Its geographic scope is explicitly global. The dataset is built from 233 uniformly distributed sampling points across all continents and includes urban built-up areas, rural settlements, mountainous terrain, plains and basins, and desert and water-body regions. This scene diversity is central to the dataset’s role as a registration benchmark because multimodal matching is reported to be challenged by geometric, radiometric, and viewpoint discrepancies across modalities (Wu et al., 20 Mar 2025).

The raw imagery is substantially larger than the benchmark tiles used for matching. Original image resolutions span 3 m to 500 m per pixel, with pixel dimensions ranging from $7{,}000\times5{,}000$ to $20{,}000\times15{,}000$. These large originals are partitioned into uniform $512\times512$ sub-images for downstream matching. Initial image-pair generation produced 170,162 map–visible pairs; after quality screening, 121,781 valid aligned pairs remained (Wu et al., 20 Mar 2025).

A plausible implication is that MapData is intended not merely as a collection of map images, but as a large-scale geometric correspondence resource in which cross-modal alignment quality is the primary annotation target.

2. Data construction and alignment pipeline

MapData’s construction pipeline combines manual control-point annotation with automated refinement. The reported workflow begins with a coarse alignment stage in which 200 well-distributed reference point pairs $\{p_i \leftrightarrow p'_i\}_{i=1}^{200}$ are manually annotated. An initial homography $H\in\mathbb{R}^{3\times3}$ is then estimated via RANSAC with maximum reprojection error $\le 3$ px (Wu et al., 20 Mar 2025).

After coarse registration, both large images are partitioned into a non-overlapping grid of $512\times512$ patches. Grid corners are projected by $H$ to form roughly aligned sub-images. The fine alignment stage then applies iterative template matching to refine $512\times512$0 to sub-pixel accuracy, reported as 1 px. Each refined homography is validated via the a-contrario test, and the workflow includes manual rechecking with optional addition of extra correspondences before re-solving $512\times512$1 (Wu et al., 20 Mar 2025).

The homography satisfies

$512\times512$2

and is recovered by minimizing the symmetric reprojection error under RANSAC. This makes MapData structurally different from datasets that provide sparse keypoint lists or semantic masks; its annotation target is a single projective transform per pair rather than direct semantic region labels (Wu et al., 20 Mar 2025).

The annotation workflow is described as hybrid manual–automated. It consists of manual coarse control-point labeling, automated patch partitioning and template matching for homography refinement, and then a-contrario validation plus manual re-inspection. The final annotation output for each $512\times512$3 pair is one ground-truth homography matrix $512\times512$4, from which dense correspondences are implicit (Wu et al., 20 Mar 2025).

3. Data organization, statistics, and annotation semantics

After filtering, MapData contains 121,781 aligned pairs. The prescribed split is 109,871 training pairs, 10,000 validation pairs, and 1,910 test pairs (Wu et al., 20 Mar 2025).

The annotation semantics are deliberately geometric. Each image pair receives one homography matrix, and no explicit keypoint-pair list or semantic masks are provided. This design places MapData closer to geometric registration benchmarks than to datasets centered on semantic segmentation, OCR, or cartographic content labeling (Wu et al., 20 Mar 2025).

The following table summarizes the core published dataset statistics.

Component	Value
Sampling points	233
Initial map–visible pairs	170,162
Valid aligned pairs	121,781
Tile size	$512\times512$5
Train / Val / Test	109,871 / 10,000 / 1,910

The scene distribution is described as balanced by region and land-cover type, but exact per-category counts are not published. The dataset nonetheless explicitly spans five scene types: urban built-up areas, rural settlements, mountainous terrain, plains and basins, and desert and water-body regions (Wu et al., 20 Mar 2025).

This annotation design distinguishes MapData from several adjacent map-data resources. MapReader centers on scanned or born-digital historical maps and patch-level semantic classification over approximately 16K nineteenth-century Ordnance Survey sheets and approximately 30.5M patches (Hosseini et al., 2021). CartoMark organizes roughly 10,400 qualified original maps into benchmarks for text annotation recognition, scene classification, super-resolution reconstruction, and style transfer (Zhou et al., 2023). MapQA and MapIQ instead target map question answering rather than geometric registration (Chang et al., 2022, Srivastava et al., 15 Jul 2025). These comparisons suggest that MapData occupies a specialized niche: multimodal geometric alignment between electronic maps and visible imagery.

4. Benchmark protocol and evaluation regime

MapData-test evaluates matching robustness by simulating three difficulty levels—Easy, Normal, and Hard—through random rotation, scaling, and translation perturbations. The reported perturbation ranges are rotation of $512\times512$6, scaling of $512\times512$7, and translation of $512\times512$8–$512\times512$9. Estimated homographies are computed with RANSAC using a 1.5 px inlier threshold, 10,000 iterations, and 0.9999 confidence (Wu et al., 20 Mar 2025).

Performance is reported using the AUC of corner-reprojection error at thresholds of 3, 5, and 10 pixels. The published sample benchmark table reports AUC@5 px percentages for a range of traditional and deep methods, including RIFT, HOWP, POS-GIFT, SuperGlue, LightGlue, RoMa, MINIMARoMa, FastMapGlue, and MapGlue (Wu et al., 20 Mar 2025).

Method	Easy	Normal	Hard
RIFT [28]	0.75	0.16	0.31
HOWP [32]	7.83	2.02	1.23
POS-GIFT [35]	15.82	10.37	5.45
SuperGlue [40]	2.27	0.83	0.20
LightGlue [41]	0.26	0.17	0.08
RoMa [50]	2.49	1.17	0.37
MINIMARoMa [54]	24.96	10.78	3.26
FastMapGlue	54.65	43.81	27.64
MapGlue	51.71	42.71	34.60

These results show two reported patterns. First, multimodal map–visible matching on MapData is substantially harder than standard unimodal feature matching, as reflected by the weak AUC@5 px scores of several general-purpose baselines. Second, the stronger methods retain relatively high performance under the Hard setting, with MapGlue’s reported 34.60 AUC@5 px on Hard exceeding FastMapGlue’s 27.64 in that regime, while FastMapGlue is higher on Easy and Normal (Wu et al., 20 Mar 2025).

A plausible implication is that MapData stresses robustness to severe cross-modal distortions rather than rewarding only high performance in mild registration conditions.

5. MapGlue and the role of semantic context

MapData is introduced together with MapGlue, a universal multimodal remote-sensing image matching framework. MapGlue is described as integrating semantic context with a dual graph-guided mechanism to extract cross-modal invariant features, enabling global-to-local interaction and enhancing descriptor robustness against modality-specific distortions (Wu et al., 20 Mar 2025).

The benchmark results above are therefore not incidental to the dataset: they are part of a co-designed research program in which MapData provides the scale and annotation regime needed to evaluate a semantics-driven matcher. The paper further reports that MapGlue generalizes effectively to unseen modalities without retraining and shows strong generalization capabilities on other modality matching tasks for which it was not specifically trained (Wu et al., 20 Mar 2025).

MapData also supports tests beyond its own held-out split. The associated work reports evaluation on five public multimodal datasets—SRIF, MultiResSAR-Low, CoFSM, DIODE, and LGHD LWIR/RGB—alongside the in-domain benchmark on MapData-test. This suggests that the dataset is positioned both as a standalone benchmark and as a training substrate for broader multimodal matching research (Wu et al., 20 Mar 2025).

A related connection appears in Plan2Map, where map registration in planning documents uses MINIMA-LoFTR during sliding-window search over basemap alignments (Degen et al., 1 Jun 2026). Although Plan2Map is a document-grounded geospatial reconstruction benchmark rather than a map–visible registration dataset, this overlap suggests that cross-modal alignment modules similar in spirit to those benchmarked on MapData can be relevant in downstream geospatial workflows.

6. Position within the broader map-data benchmark ecosystem

Within the broader literature represented here, MapData belongs to a family of map-centric datasets, but its target task is distinct. Other resources address different technical strata of map understanding.

MapReader provides a Python pipeline for scanned or born-digital map collections, including retrieval, preprocessing, patch extraction, interactive annotation, deep model training, and GeoJSON or CSV export. Its case study focuses on rail infrastructure and buildings on nineteenth-century Ordnance Survey sheets, rather than cross-modal registration (Hosseini et al., 2021). CartoMark organizes map images for text annotation recognition, scene classification, super-resolution reconstruction, and style transfer, not homography recovery (Zhou et al., 2023). CMEdataset focuses on detecting problematic maps of China via bounding boxes, polygons, and segmentation masks across five error categories (Xu et al., 10 Apr 2025).

MapQA and MapIQ formalize visual question answering over thematic maps. MapQA contains approximately 800K question-answer pairs over approximately 60K choropleth images and introduces V-MODEQA for image-to-table extraction followed by reasoning (Chang et al., 2022). MapIQ comprises 14,706 question-answer pairs over choropleths, cartograms, and proportional symbol maps, and evaluates multimodal LLMs across six visual analytical tasks (Srivastava et al., 15 Jul 2025). These datasets measure interpretive reasoning over map graphics rather than registration accuracy.

Plan2Map and MapTrace occupy yet other positions. Plan2Map reconstructs geospatial boundaries from planning documents and evaluates GeoPlanAgent with IoU-based geospatial boundary scoring (Degen et al., 1 Jun 2026). MapTrace generates synthetic route-tracing data over 4,000 maps and 23,000 path annotations, targeting fine-grained path following rather than map–visible homography estimation (Panagopoulou et al., 22 Dec 2025).

This comparative picture suggests that “map data” in current arXiv literature is not a single task category. It includes at least semantic extraction, question answering, compliance detection, route tracing, document-grounded reconstruction, and multimodal registration. MapData’s contribution is specifically to the registration branch, where the ground truth is a homography and the operative challenge is cross-modal geometric correspondence.

7. Significance, limitations, and research implications

MapData’s significance lies in the conjunction of three published properties: global geographic coverage, scale sufficient for deep learning, and rigorous homography-based ground truth. The MapGlue paper explicitly characterizes it as addressing the scarcity of scalable multimodal benchmarks and describes it as the largest publicly released electronic-map–visible imagery dataset for multimodal remote-sensing registration to date (Wu et al., 20 Mar 2025).

Its methodological strengths follow directly from the construction protocol. Manual control points, iterative template matching, a-contrario validation, and manual re-inspection together define a relatively strict alignment workflow. The use of a single homography per tile also makes evaluation crisp and computationally standardized (Wu et al., 20 Mar 2025).

At the same time, the dataset’s scope imposes clear boundaries. The annotation output is limited to one homography per pair; there are no explicit semantic masks, no keypoint-pair lists, and no published exact per-category scene counts. The benchmark evaluates registration under synthetic perturbations, which is well suited to controlled comparison but does not by itself exhaust the space of real-world cartographic distortions (Wu et al., 20 Mar 2025). This suggests that MapData is best understood as a foundational registration benchmark rather than a complete representation of all map-understanding problems.

A plausible implication is that future work may combine MapData-like geometric supervision with datasets from neighboring subfields. For example, semantic priors from MapReader (Hosseini et al., 2021), VQA-style reasoning structures from MapQA or MapIQ (Chang et al., 2022, Srivastava et al., 15 Jul 2025), or large-scale graph alignment ideas from UM3 (Ying et al., 23 Aug 2025) could be integrated into broader multimodal map systems. Such an overview would extend beyond the published claims, but it is consistent with the present fragmentation of map-data research into complementary benchmarks and task formulations.