Warped Alignment of Reprojected Graphs (WARG)

• WARG is a family of geometry-aware alignment techniques that preserve intrinsic data structures by avoiding naive flattening.
• It combines unified graph learning with explicit 3D reprojection and candidate matching for accurate cross-view lunar rover localization.
• The framework extends to manifold and attributed graph alignment, offering robust performance in warped image matching and metric warping scenarios.

Searching arXiv for WARG and related papers to ground the article in current literature. Warped Alignment of Reprojected Graphs (WARG) denotes a family of geometry-aware alignment procedures in which observations are first represented in a native geometric domain and then matched under an explicit warp model. In its explicit usage, WARG is a framework for cross-view lunar rover localization that combines unified graph learning with reprojected graph matching under candidate $SE(3)$ transforms (Chen et al., 9 Jun 2026). In a broader conceptual sense, the same phrase describes manifold-aware alignment of geodesic graphs induced by warped images (Sharma et al., 2020) and decoder-level metric warping that aligns an attribute-manifold diffusion kernel to graph geometry (Labarthe et al., 30 Jan 2026). Across these settings, the common pattern is to avoid naive flattening or single-space embedding and instead preserve the geometry in which the data are naturally defined.

1. Terminological scope and conceptual boundaries

The acronym WARG appears explicitly in the lunar localization framework "Globally Localizing Lunar Rover in Pixels via Graph Alignment" (Chen et al., 9 Jun 2026). By contrast, the acronym does not appear in the 2020 paper "Simplicial Complex based Point Correspondence between Images warped onto Manifolds" (Sharma et al., 2020), nor in the 2026 paper "Aligning the Unseen in Attributed Graphs: Interplay between Graph Geometry and Node Attributes Manifold" (Labarthe et al., 30 Jan 2026). In both of those cases, the supplied formulation maps their methods to WARG conceptually rather than terminologically.

This distinction matters because WARG is not a single standardized algorithm across all graph alignment literature. In the lunar setting, it denotes a concrete end-to-end localization framework. In the warped-image setting, it denotes alignment of reprojected graphs on manifolds through graph-induced simplicial complexes and a constrained quadratic assignment problem. In the attributed-graph setting, it denotes a two-phase geometric alignment process in which a graph is reprojected onto an attribute manifold through a diffusion kernel, and the manifold metric is then warped to align with graph structure.

A common misconception is therefore to treat WARG as synonymous with a learned assignment matrix or with graph neural network message passing. The explicit lunar WARG uses neither: it employs no learned assignment matrix, and "graph learning" refers to saliency-driven node selection, shared descriptors across views and scales, and structural consistency induced by coherent reprojection under a single candidate transform (Chen et al., 9 Jun 2026).

2. Cross-view lunar localization formulation

In its explicit form, WARG addresses cross-view lunar localization: given a rover-view image $I_r$ and a satellite-view image $I_s$, the task is to estimate the rover’s absolute pose in a global selenocentric frame by matching $I_r$ to $I_s$, thereby avoiding the drift accumulation of dead-reckoning and visual odometry in GNSS-denied environments (Chen et al., 9 Jun 2026). The method is designed around three lunar-specific challenges: inter-entity entanglement, inter-viewpoint divergence, and simulation-to-real domain shift.

The framework combines two components. The first is unified graph learning with shared parameters across views, intended to produce viewpoint-invariant features and salient nodes. The second is reprojected graph matching, in which the rover graph is warped into the satellite image under candidate $SE(3)$ transforms and scored by feature similarity with joint saliency weighting (Chen et al., 9 Jun 2026).

The end-to-end pipeline is fixed by the paper. A frozen DINOv3 encoder and a shared DPT decoder produce a multi-scale feature pyramid for both $I_r$ and $I_s$. On the rover side, saliency and feature-refinement heads select top-$N$ salient nodes across scales, attach refined descriptors, and define edges by rover-frame $3$D displacements using depth. On the satellite side, the same shared heads produce dense feature and saliency maps at all scales. Reprojected graph alignment then evaluates candidate rover translations around an initial prior by reprojecting rover nodes into $I_r$0, sampling satellite features at the resulting coordinates, aggregating multi-scale saliency-weighted cosine similarities, and selecting the maximum a posteriori hypothesis (Chen et al., 9 Jun 2026).

The rover graph is

$I_r$1

where node $I_r$2 has pixel coordinate $I_r$3, refined descriptor $I_r$4, saliency weight $I_r$5, and rover-frame point $I_r$6 derived from depth. Edges are undirected and encode relative $I_r$7D displacement,

$I_r$8

The satellite side is kept dense in practice rather than sparsified into an explicit graph because the satellite image covers a larger area (Chen et al., 9 Jun 2026).

The multi-scale feature pyramid uses strides

$I_r$9

At each scale $I_s$0, the shared heads produce a saliency map

$I_s$1

and a refined descriptor map

$I_s$2

Rover nodes are selected by top-$I_s$3 saliency scores at each scale, with $I_s$4, and are lifted to $I_s$5D by

$I_s$6

3. Reprojection, alignment objective, and inference mechanics

The geometric core of WARG is full $I_s$7D reprojection with depth rather than a planar homography approximation. Let the selenocentric world frame be $I_s$8, the satellite camera be $I_s$9, and the rover camera be $I_r$0. Intrinsics $I_r$1 and $I_r$2, rover rotation $I_r$3, rover attitude, and satellite parameters are treated as known, while the method estimates rover translation $I_r$4 over a sampled search region (Chen et al., 9 Jun 2026).

For candidate translation $I_r$5, the world point of rover node $I_r$6 is

$I_r$7

Satellite projection is then

$I_r$8

with

$I_r$9

Equivalently,

$I_s$0

The paper states a planar special case for completeness but does not use it; WARG instead uses full $I_s$1D reprojection with depth, avoiding planarity assumptions (Chen et al., 9 Jun 2026).

For each candidate $I_s$2, WARG samples a satellite descriptor and saliency at the reprojected location:

$I_s$3

Joint saliency weighting is defined by

$I_s$4

$I_s$5

Cross-view similarity is cosine similarity,

$I_s$6

and the candidate logit is

$I_s$7

A softmax over candidates yields

$I_s$8

with the MAP candidate used as the localization output (Chen et al., 9 Jun 2026).

Training uses only the negative log-likelihood over the nearest-to-ground-truth candidate,

$I_s$9

The trainable components are the DPT decoder and the two MLP heads $SE(3)$0 and $SE(3)$1; the DINOv3 backbone is frozen. No supervised correspondences or pose regression are used, and there is no learned assignment matrix. Structural consistency is enforced implicitly by coherent reprojection of all rover nodes under a single candidate translation together with the joint saliency mechanism (Chen et al., 9 Jun 2026).

The inference procedure consists of shared feature extraction, shared saliency and descriptor prediction, rover graph construction, candidate sampling, candidate-wise reprojected matching, and selection of

$SE(3)$2

An optional finer resampling around $SE(3)$3 may be used for local refinement, although the paper’s core method uses discrete candidate selection (Chen et al., 9 Jun 2026).

4. Higher-order manifold alignment on warped images

A conceptually earlier realization of warped alignment of reprojected graphs appears in "Simplicial Complex based Point Correspondence between Images warped onto Manifolds" (Sharma et al., 2020). The motivating setting is image matching when landmarks naturally lie on a curved $SE(3)$4D manifold $SE(3)$5 embedded in $SE(3)$6, including spherical, cylindrical, elliptic, and conic imaging geometries. Distances and angles are measured intrinsically through geodesics on $SE(3)$7, not through Euclidean straight lines in the image plane.

The central motivation is that flattening to the plane introduces nonuniform scale, angle, and area distortions that corrupt pairwise and local higher-order relations. On the sphere $SE(3)$8, for example, an equirectangular map scales longitudinal distances by $SE(3)$9 and stretches areas near the poles. The method therefore formulates matching as finding a bijective map between two graph-induced simplicial complexes built directly from manifold landmarks (Sharma et al., 2020).

Starting with landmarks $I_r$0 on $I_r$1, a geodesic metric $I_r$2 is used to construct an undirected graph

$I_r$3

by connecting $I_r$4 whenever their geodesic $I_r$5-neighborhoods intersect:

$I_r$6

The graph-induced simplicial complex $I_r$7 contains one $I_r$8-simplex for every $I_r$9-clique in $I_s$0. Its $I_s$1-skeleton $I_s$2 contains all simplices of dimension at most $I_s$3. This replaces pairwise graph structure by a higher-order complex of vertices, edges, triangles, and higher cliques (Sharma et al., 2020).

Each $I_s$4-skeleton is represented by a boundary matrix

$I_s$5

whose rows index $I_s$6-simplices and columns index $I_s$7-simplices, with entry

$I_s$8

iff the $I_s$9-simplex is a facet of the $N$0-simplex. Two simplices are adjacent if they share a face, and each simplex is described through the barycenters of simplices in its neighborhood. For a simplex $N$1, the method computes an affine weight vector

$N$2

by least squares subject to

$N$3

This manifold-aware descriptor compactly encodes local geodesic neighborhood geometry (Sharma et al., 2020).

Matching is posed as a constrained quadratic assignment problem over successive skeletons. If simplex counts match, the assignment matrix is

$N$4

with bijectivity constraints

$N$5

Descriptor costs are

$N$6

$N$7

With a block-structured geodesic-cost matrix $N$8, the objective is

$N$9

subject to the bijectivity constraints above (Sharma et al., 2020).

The algorithm proceeds top-down from the highest skeleton dimension to the lowest. For each $3$0, it assembles boundary matrices, computes adjacency and neighborhoods, derives affine weight vectors, constructs the cost matrices, solves the QAP by spectral relaxation using the principal eigenvector of $3$1, projects to the Birkhoff polytope via the Hungarian algorithm, and propagates confident higher-dimensional matches to lower-dimensional facets. The final output is $3$2, the vertex-level correspondence (Sharma et al., 2020).

On $3$3, the intrinsic metric is

$3$4

with optional spherical invariants such as angles, triangle area

$3$5

and orientation

$3$6

The core pipeline, however, uses the manifold-aware $3$7 descriptors built from geodesic neighborhoods (Sharma et al., 2020).

5. Metric warping and reprojection on attributed graphs

A different but formally related use of the same conceptual pattern appears in "Aligning the Unseen in Attributed Graphs: Interplay between Graph Geometry and Node Attributes Manifold" (Labarthe et al., 30 Jan 2026). Here the object is an undirected weighted attributed graph

$3$8

with adjacency $3$9 and node attributes $I_r$00. The method argues that the standard approach of simultaneously reconstructing node attributes and graph structure is geometrically flawed because it merges potentially incompatible metric spaces.

The method therefore separates manifold learning from structural alignment. In Phase 1, a custom VAE learns the intrinsic attribute manifold $I_r$01 from $I_r$02. The encoder is

$I_r$03

and the decoder is

$I_r$04

The decoder induces a pullback Riemannian metric

$I_r$05

and latent geodesic distances $I_r$06 are approximated either by a differentiable grid-and-shortest-path scheme or by a differentiable linear-segment line integral (Labarthe et al., 30 Jan 2026).

The reprojection step is the manifold diffusion kernel

$I_r$07

with

$I_r$08

This operator maps fixed latent coordinates to an $I_r$09 kernel matrix that can be aligned directly to graph structure. The multi-scale set of diffusion times is chosen from the graph spectrum:

$I_r$10

using $I_r$11 log-spaced times in $I_r$12 (Labarthe et al., 30 Jan 2026).

Phase 2 freezes the encoder and warps the decoder metric so that the manifold diffusion kernel better matches the graph. The alignment objective is

$I_r$13

Warping is implemented solely through $I_r$14; the latent coordinates $I_r$15 cannot move. The method interprets the resulting change in geodesics as a structural descriptor of incompatibility between attribute and graph geometries (Labarthe et al., 30 Jan 2026).

Pairwise distortion is measured by

$I_r$16

where superscripts $I_r$17 and $I_r$18 denote the pre-alignment and post-alignment manifolds. A robust standardized score is

$I_r$19

with $I_r$20, and node-level distortion is

$I_r$21

High positive $I_r$22 or large $I_r$23 indicates large warping required to reconcile attribute manifold and observed connectivity; negative values indicate neighborhoods well explained by attribute proximity (Labarthe et al., 30 Jan 2026).

This formulation is conceptually close to WARG because reprojection is performed through a geometry-induced kernel and alignment is achieved by explicit metric deformation rather than by collapsing the two spaces into a single latent geometry. The paper characterizes this as a known-correspondence metric alignment distinct from Gromov–Wasserstein coupling, since the identification $I_r$24 is fixed (Labarthe et al., 30 Jan 2026).

6. Empirical characteristics, assumptions, and significance

The explicit lunar WARG reports an average test localization error of $I_r$25 on the synthetic LuSNAR dataset, $I_r$26 in zero-shot transfer to the synthetic lunar south pole region, and $I_r$27 on real YuTu-2 data for the panoramic variant WARG-P within a $I_r$28 search area. With satellite resolution $I_r$29, this corresponds to nearly one-pixel precision. The model contains $I_r$30 trainable parameters, corresponding to $I_r$31 of previous lightweight baselines, and runs at $I_r$32 on an NVIDIA RTX A6000 GPU (Chen et al., 9 Jun 2026).

The ablations in the same paper indicate that increasing nodes per scale improves robustness and plateaus around $I_r$33 nodes per scale, that independent weights catastrophically fail on the South split with mean error $I_r$34, that precision scales with resolution while remaining stable down to $I_r$35 px, and that enlarging the search region from $I_r$36 to $I_r$37 increases error only marginally. The paper also reports tolerance to motion blur and occlusion and describes emergent spatial awareness in the learned saliency and similarity maps, including concentration on crater rims and boulders and transfer to rover-to-rover correspondence (Chen et al., 9 Jun 2026).

The manifold simplicial-complex method reports very low error on spherical and warped-image benchmarks, including $I_r$38 on Kamaishi, $I_r$39 on Desktop, $I_r$40 on Parking, and $I_r$41 on Table, while planar or descriptor-based alternatives on the same warped data often show large errors. On flattened spherical images it remains substantially more accurate, for example $I_r$42 on Desktop_flat and $I_r$43 on Table_flat, and the paper reports up to $I_r$44 reduction in matching error on warped images and up to $I_r$45 reduction on flat images relative to state-of-the-art baselines. The method is also described as robust under rotations, reflections, scaling, shear, point missingness, and two random noise models (Sharma et al., 2020).

The attributed-graph alignment method reports, on the synthetic Swiss-roll experiment, $I_r$46 scores of $I_r$47 and $I_r$48 for two geodesic variants, ROC AUC values of $I_r$49 and $I_r$50, and ARI of $I_r$51 for identifying perturbed connectivity. On the Île-de-France experiment, it reports that regressing transport convenience on attributes or pairwise similarities yields $I_r$52, supporting the claim that the misalignment signal is not contained in attributes alone but emerges from their conflict with topology (Labarthe et al., 30 Jan 2026).

Across all three formulations, the principal assumptions are explicit. Lunar WARG assumes known camera calibration and rover attitude, a reasonable translation prior, and depth estimates; large depth errors can degrade reprojection accuracy, and little or no informative overlap leaves alignment underconstrained (Chen et al., 9 Jun 2026). The simplicial-complex method assumes known or estimable manifold geometry and sufficient sampling density for higher-order cliques; incorrect manifold models, highly nonisometric settings, extreme sparsity, or differing topology can make geodesic neighborhoods or consistency constraints misleading (Sharma et al., 2020). The attributed-graph method is sensitive to ill-conditioned spectra, widespread attribute-structure mismatch, and attribute scaling; the paper recommends monitoring average $I_r$53 and using standardized attributes (Labarthe et al., 30 Jan 2026).

Taken together, these works define WARG less as a single algorithm than as a geometric alignment principle: represent observations in the space where their structure is intrinsic, reproject them into a comparable form, and perform alignment through warp-aware constraints rather than through flattening, naive Euclideanization, or unconstrained joint embeddings. In the lunar setting this yields drift-free global localization in GNSS-denied environments; in warped imaging it yields robust correspondence on manifolds; and in attributed graphs it turns metric conflict into an interpretable descriptor of heterophily, shortcuts, and anomalies (Chen et al., 9 Jun 2026).