GlueMap: Multidisciplinary Gluing for Mapping and Analysis

• GlueMap is a family of methodologies that use gluing techniques to integrate local data into coherent global structures across multiple domains.
• The approach combines classical geometric and deep feedforward inference in SfM, semantic context in multimodal remote sensing, and algebraic strategies in Floer theory.
• Practically, GlueMap enhances 3D reconstruction accuracy, robust cross-modal image matching, and theoretical assembly in algebraic topology through modular pipeline designs.

GlueMap denotes a family of methodologies and systems—distinct in context but unified by the central role of "gluing" for structure assembly, matching, or mapping across multiple data modalities or domains. Three principal research lines use the term: (1) GlueMap for structure-from-motion (SfM), (2) MapGlue for multimodal remote-sensing image matching, and (3) gluing maps in (algebraic/topological) Floer theory. Each leverages gluing—whether geometric, algebraic, or algorithmic—as a tool for integrating fragmented local inferences or structures into global or cross-modal solutions.

1. GlueMap in Structure-from-Motion Pipelines

GlueMap is a modular SfM system explicitly designed to integrate classical geometric optimization with modern feedforward deep inference, capitalizing on their complementary strengths for robust 3D reconstruction in challenging scenarios (Pan et al., 25 May 2026).

Pipeline Overview

The GlueMap pipeline comprises four stages:

1. View Graph Initialization: A learned image retrieval scheme (e.g., SALAD) identifies candidate neighbors for each image. These are filtered by a Doppelganger (DG) disambiguation network, which scores each candidate pair to suppress “lookalikes.” The system adaptively forms a sparsely connected yet complete graph by thresholding DG scores while dynamically lowering the threshold to ensure global connectivity.
2. Feedforward Local Inference: For every image, its local 1-hop "star" (itself and its immediate neighbors) is processed in parallel by the π³ (Pi³) multi-view transformer. π³ predicts per-image relative poses, depth maps, focal lengths, and soft correspondence tracks, which are merged across overlapping stars and snapped to classical SIFT features.
3. Global Motion Averaging: Intrinsics (focal lengths) are aggregated by median across stars. Rotations are globally averaged by minimizing a robust Huberified SO(3) geodesic loss over the view graph, weighted by overlap ratios derived from forward-backward reprojections. A similarity averaging step aligns local scale and translation, rescaling all predicted depths.
4. Augmented Bundle Adjustment (A-BA): Joint optimization over camera parameters and 3D structure leverages three track types: SIFT (classical), deep (from π³), and virtual (reprojected from depths). SIFT tracks are prioritized; deep tracks fill in when SIFT correspondences are sparse, and virtual points are used as additional constraints.

Mathematical Objectives

Key loss components:

• Reprojection Consistency: Forward-backward error between local depth-induced 3D points and their re-projections across image pairs.
• Rotation Averaging: Robust global alignment of camera orientations: $$E_{rot}(R) = \sum_{(i,j)\in E} \rho(o_{ij}\,\angle(R_j R_i^T, R_{ij}^l))$$ where $\rho$ denotes the Huber loss and $o_{ij}$ the overlap ratio.
• Similarity Averaging: Alignment of camera centers and per-star scale:

$$E_{sim}(c,s)=\sum_{l,(i,j)\in S_l} o_{ij}^l\,\|R_{ij}^{l^T}\,t_{ij}^l - s_l\,(c_i-c_j)\|^2$$

• Bundle Adjustment Error: Robust reprojection loss over all tracks (SIFT, deep, and virtual).

Network and Fusion

π³ functions as a multi-view transformer, encoding images as patch tokens and using self- and cross-attention to propagate and refine information intra- and inter-view. The outputs directly inform global optimization: relative pose/scale alignments and soft correspondences feed into the motion averaging and A-BA stages.

Quantitative Performance

Across multiple challenging datasets, GlueMap exhibits state-of-the-art AUC for pose recovery in both typical and “hard” (low-texture, small overlap, symmetry) scenarios. For ETH3D, AUC@1° reaches 53.0 (vs. SIFT 45.6, π³ 13.2); for SMERF minimal-overlap benchmarks, GlueMap maintains reasonable accuracy (AUC@20°=82.0, where all other methods fall below 20). On large-scale indoor/outdoor scenes in LaMAR, only GlueMap achieves robust reconstructions when other feedforward or classical pipelines fail (Pan et al., 25 May 2026).

2. MapGlue for Multimodal Remote-Sensing Image Matching

MapGlue is an advanced, semantics-driven pipeline for multimodal remote sensing image (MRSI) matching, designed to address severe geometric, photometric, and viewpoint variances across disparate sensing modalities. It is jointly introduced with MapData, a large-scale, global MRSI dataset (Wu et al., 20 Mar 2025).

System Architecture

MapGlue operates in two principal stages:

1. Semantic Context Embedding:
   • Saliency-aware keypoints and structural descriptors are extracted using a modified SuperPoint (referred to as SES).
   • Fine-grained semantic descriptors are derived from the MobileSAM encoder.
   • A multi-layer perceptron fuses these into unified “saliency descriptors.”
2. Dual Graph-Guided Matching:
   • A stack of transformer layers operates over “intra-image” undirected dynamic sparse graphs (self-attention) and “inter-image” directed semantic graphs (cross-attention).
   • The intra-image graph adaptively progresses from global to local context, governed by decreasing connection radii $\varepsilon^{(l)}$.
   • The inter-image graph uses top-K cosine-similar features as directed edges, focusing attention on genuinely cross-modal-invariant structures.

Mathematical Details

• Feature Extraction: For each keypoint $p_i$ in modality $m$:

$$f_i^m = \Phi_{\rm str}(I^m)|_{p_i},\quad s_i^m = \Phi_{\rm sem}(I^m)|_{p_i},\quad d_i^m = \mathrm{MLP}[f_i^m \| s_i^m]$$

• Graph-Guided Attention: Self-attention is performed via normalized adjacency, cross-attention via top-K cosine similarity neighbors.
• Loss Function: A quadruplet contrastive-style cross-entropy blends positive, negative, false-positive, and false-negative correspondences, optionally supplemented by a margin triplet loss.

MapData: Dataset

• Scale and Diversity: 233 global sampling points; ~122k aligned map–visible image pairs (cropped to $512 \times 512$).
• Ground Truth: Two-stage process—manual tie-point annotation and RANSAC homography, followed by iterative refinement via template matching and a-contrario statistical validation.
• Splits: 109,871 training, 10,000 validation, 1,910 test pairs.

Experimental Results

• MapGlue achieves the highest accuracy (AUC@5px) across Easy/Normal/Hard protocols versus all public baselines (see Table 1).
• Generalization: MapGlue maintains strong performance on multiple unseen datasets, outperforming prior state-of-the-art by large multiplicative margins (e.g., +81.3% to +173.7% on SRIF).

Method	Easy	Normal	Hard
SuperGlue	2.27	0.83	0.20
XoFTR	20.79	9.20	2.39
RoMa	24.96	10.78	3.26
MINIMARoMa	25.59	11.77	3.13
FastMapGlue	54.65	43.81	27.64
MapGlue	51.71	42.71	34.60

Table 1. MapData-test AUC@5 px for Easy/Normal/Hard (Wu et al., 20 Mar 2025).

3. Gluing Maps in Floer Homology and Algebraic Geometry

The concept of a "gluing map" has deep roots in algebraic topology and algebraic geometry, particularly in the construction of invariants via excision, degeneration, or matching of local data.

Sutured Monopole Floer Homology

• For a balanced sutured manifold $(M, \Gamma)$, gluing maps (denoted $\rho$0) enable the assembly of the monopole Floer group $\rho$1 from those of sutured submanifolds. The construction proceeds via closure data, cobordism (Floer excision), and canonical isomorphism on local systems.
• Fundamental properties:
  • Functoriality: successive gluings commute up to a unit in the Novikov ring.
  • Naturality: the maps are independent of choices (e.g., decomposition, contact structure) up to a multiplicative unit.
  • Cobordism: The gluing map forms a building block for cobordism maps in the category of sutured manifolds (Li, 2018).

Gluing Formalism for Punctured Logarithmic Maps

• In logarithmic Gromov–Witten theory, gluing formulas reconstruct the moduli and (virtual) fundamental classes of stable log maps by assembling data from splittings at edges in their tropicalizations.
• The master formula equates the virtual class of a glued moduli space to a weighted sum (by tropical multiplicity $\rho$2) of the product of classes over the split pieces:

$\rho$3

• This approach unifies and generalizes classical degeneration formulas (Li–Ruan, Jun Li) and underpins the canonical wall structure for K3 surfaces via wall-scattering products (Gross, 2023).

4. Algorithmic and Practical Considerations

SfM/Matching Contexts

• Scalability: GlueMap in SfM handles up to thousands of images by leveraging star-shaped local subgraphs (bounded star size, GPU-friendly).
• Fault Tolerance: Robustness emerges from adaptive view graph construction (DG filtering for lookalike disambiguation), fusion of heterogeneous tracks, and robustified global/BA objectives.
• Implementation: Critical engineering choices include neighborhood sizes ($\rho$4), snapping radius ($\rho$5 px), and edge filtering thresholds ($\rho$6 px).

Limitations and Ongoing Developments

• Distributional Limits: GlueMap's performance relies on the training regime of its feedforward modules (currently pinhole camera geometry only).
• Motion Constraints: Pure rotations/constrained motion sequences challenge the depth-based prior; enhancements require explicit rotational priors or camera models.
• Dataset Domain: MapGlue is explicitly cross-modal for remote sensing but is not a direct substitute for geometric-matching in consumer vision tasks.

Open-Source Availability

GlueMap (SfM) and MapGlue (MRSI) are available for benchmarking and extension at their respective repositories:

• SfM: https://github.com/colmap/gluemap (Pan et al., 25 May 2026)
• MRSI: https://github.com/PeihaoWu/MapGlue (Wu et al., 20 Mar 2025)

5. Broader Impact and Applications

• Remote Sensing: Accurate cross-modal alignment is essential for geolocalization, change detection, disaster response, and dynamic map updating—tasks directly supported by MapGlue and MapData's coverage of 233 globally diverse sites (Wu et al., 20 Mar 2025).
• 3D Reconstruction: GlueMap serves as a robust backbone for large-scale, heterogeneous image-based modeling, handling minimal overlap, symmetry, and ambiguous context.
• Theoretical Insights: The gluing map formalism in algebraic geometry and Floer theory provides a unifying language for assembling global structures from local data, underlying modern developments in Gromov–Witten theory and topological invariants (Gross, 2023, Li, 2018).

A plausible implication is that continued advances in gluing-based methodologies—and their increasingly open, modular implementations—will further diminish the divide between local, learned inference and global, structure-based composition. This trend is evident both in hybrid SfM pipelines and in the assembly of moduli or topological invariants from local geometric input.