MatchAnything-ELoFTR: Pretrained SAR-Optical Matcher

• The paper demonstrates that MatchAnything-ELoFTR, pretrained on synthetic cross-modal pairs, attains approximately 3.4 px mean error and 0% failure in zero-shot SAR-optical registration.
• It employs the EfficientLoFTR architecture with a frozen DINOv2 feature extractor and a coarse-to-fine transformer head to compute dense correspondences.
• The evaluation shows that deployment protocol choices, such as normalization, tile size, and RANSAC filtering, substantially impact registration performance.

Searching arXiv for MatchAnything and related matcher papers to ground the article with primary sources. MatchAnything-ELoFTR, often abbreviated as MA-ELoFTR in evaluation tables, is a pretrained cross-modal image matcher built on the EfficientLoFTR backbone and assessed as a fixed off-the-shelf artifact for zero-shot SAR–optical satellite registration in "Are Pretrained Image Matchers Good Enough for SAR-Optical Satellite Registration?" (Corley et al., 11 Apr 2026). In that study it appears as one of 24 matcher families harvested via the vismatch library and is distinguished from vanilla ELoFTR not by a new inference architecture but by a different pretraining regime: further training under the MatchAnything recipe on large-scale synthetic cross-modal pairs. Within the reported SpaceNet9 evaluation, it attains a mean error of $3.4$ px with $0\%$ failure under the chosen affine-RANSAC protocol, placing it close to the best reported zero-shot results and making it a representative case in the broader question of whether foundation-model features can partially substitute for explicit cross-modal supervision (Corley et al., 11 Apr 2026).

1. Architectural identity

MatchAnything-ELoFTR is described as an EfficientLoFTR-based matcher. EfficientLoFTR, or ELoFTR, uses a frozen DINOv2 feature extractor together with a lightweight coarse-to-fine transformer head. For each image $I \in \mathbb{R}^{H \times W \times 3}$, the DINOv2 ViT backbone produces a dense feature map

$$f_{\mathrm{DINOv2}}(I) \in \mathbb{R}^{h \times w \times d},$$

after which the ELoFTR matcher pulls out two sets of paired keypoint grids, computes coarse assignment scores, and then refines them at full resolution (Corley et al., 11 Apr 2026).

The paper’s central architectural statement is that MatchAnything-ELoFTR differs from vanilla ELoFTR only in its pretraining regime. The model is therefore not introduced as a new backbone or a new matching head in the evaluation paper; rather, it is treated as a pretrained variant of ELoFTR whose distinguishing property is synthetic cross-modal pretraining. This distinction matters because the evaluation explicitly separates architectural capacity from transfer behavior. A plausible implication is that the reported performance of MA-ELoFTR should be interpreted less as evidence for a novel inference mechanism than as evidence about how pretraining data and modality exposure affect zero-shot transfer.

At the level of correspondence scoring, the paper summarizes the matcher in the following form. If $p \in \Omega_s$ denotes a pixel in the source SAR tile and $q \in \Omega_o$ denotes a pixel in the optical tile, the network produces feature vectors

$$x_p = f_{\theta_s}(I_s)_p, \qquad y_q = f_{\theta_o}(I_o)_q$$

and coarse-stage correspondence scores

$$S_{pq} = \mathrm{softmax}_q(x_p^\top \cdot y_q / \tau),$$

with analogous refinement at the fine stage (Corley et al., 11 Apr 2026).

2. Pretraining regime and the MatchAnything recipe

The MatchAnything-ELoFTR variant is reported to have been further trained under the "MatchAnything" recipe on large-scale synthetic cross-modal pairs (Corley et al., 11 Apr 2026). The evaluation paper does not reproduce the full training recipe, but it specifies the synthetic-pair generation process attributed to the MatchAnything framework of Han et al. (2024): sampling two images from different modalities or views such as optical versus SAR; applying random photometric transforms to each, including contrast, noise, and channel dropout, together with random geometric warps such as homography and thin-plate spline; and using ground-truth warp parameters to supervise a variant of the LoFTR contrastive coarse-to-fine matching loss (Corley et al., 11 Apr 2026).

The reported training objective is an InfoNCE-style cross-entropy on the ground-truth warp $\Delta$,

$L = - \sum_p \log S_{p,\,p+\Delta(p)},$

supplemented by auxiliary positional-regression penalties (Corley et al., 11 Apr 2026). Within the boundaries of the evaluation paper, these details define MatchAnything-ELoFTR as a supervised detector-free matcher whose modality bridging is induced synthetically rather than through direct adaptation to the satellite benchmarks used at test time.

This framing is important for interpreting the benchmark. The evaluation is strictly zero-shot: no fine-tuning or domain adaptation on satellite or SAR data is allowed (Corley et al., 11 Apr 2026). Consequently, the model’s behavior on SpaceNet9 and the additional cross-modal benchmarks is intended to reflect transfer from its synthetic cross-modal pretraining rather than task-specific fitting. This suggests that MA-ELoFTR occupies an intermediate position between modality-specific matchers and general-purpose foundation-feature matchers: it retains the EfficientLoFTR inference stack, but its pretraining is explicitly designed to expose the model to heterogeneous appearance transformations before deployment.

3. Zero-shot SAR–optical registration protocol

The benchmark protocol in which MatchAnything-ELoFTR is evaluated has four deterministic stages: preprocessing, tiled correspondence extraction, geometric filtering, and displacement prediction with tie-point-grounded metrics (Corley et al., 11 Apr 2026).

In preprocessing, normalization is optional and selected per model from $0\%$0. Images are then resized so that $0\%$1, where $0\%$2 (Corley et al., 11 Apr 2026). This stage is not merely cosmetic: later ablations show that some matchers are robust to normalization and others are not.

For tiled correspondence extraction, large SpaceNet9 scenes of approximately $0\%$3 px are decomposed into sliding tiles of size $0\%$4 with overlap $0\%$5. The default setting is $0\%$6 and $0\%$7, with ablations including $0\%$8 and $0\%$9. Each tile pair is passed to the matcher, and the returned keypoints are reprojected into global pixel coordinates (Corley et al., 11 Apr 2026). This tiling regime is part of the deployment protocol rather than part of the learned model, but it materially affects the outcome.

Geometric filtering aggregates all tile-level correspondences and then applies RANSAC with either an affine $I \in \mathbb{R}^{H \times W \times 3}$0 model or a homography $I \in \mathbb{R}^{H \times W \times 3}$1 model. The default reprojection threshold is $I \in \mathbb{R}^{H \times W \times 3}$2 px and the minimum inlier count is $I \in \mathbb{R}^{H \times W \times 3}$3, with ablations at $I \in \mathbb{R}^{H \times W \times 3}$4, $I \in \mathbb{R}^{H \times W \times 3}$5, and $I \in \mathbb{R}^{H \times W \times 3}$6. Pairs with fewer than $I \in \mathbb{R}^{H \times W \times 3}$7 inliers are flagged as failures, although individual tiles may still contribute to other scenes (Corley et al., 11 Apr 2026).

The final displacement estimate is evaluated by applying the estimated transform $I \in \mathbb{R}^{H \times W \times 3}$8 or $I \in \mathbb{R}^{H \times W \times 3}$9 to each ground-truth tie point $$f_{\mathrm{DINOv2}}(I) \in \mathbb{R}^{h \times w \times d},$$0 to obtain $$f_{\mathrm{DINOv2}}(I) \in \mathbb{R}^{h \times w \times d},$$1, and then computing mean pixel error:

$$f_{\mathrm{DINOv2}}(I) \in \mathbb{R}^{h \times w \times d},$$2

The benchmark also reports $$f_{\mathrm{DINOv2}}(I) \in \mathbb{R}^{h \times w \times d},$$3 for $$f_{\mathrm{DINOv2}}(I) \in \mathbb{R}^{h \times w \times d},$$4 px and the failure rate, defined as the fraction of scene pairs for which RANSAC fails entirely (Corley et al., 11 Apr 2026).

For MatchAnything-ELoFTR specifically, the best-found normalization on SpaceNet9 is z-score normalization (Corley et al., 11 Apr 2026).

4. Reported performance

On SpaceNet9 under the best-found normalization, the paper reports the following excerpted comparison among the strongest evaluated matchers (Corley et al., 11 Apr 2026).

Matcher	Mean error	Failure
XoFTR	3.0 px	0 %
RoMa	3.0 px	0 %
MINIMA-RoMa	3.4 px	0 %
MA-ELoFTR	3.4 px	0 %

For the same comparison, the paper reports $$f_{\mathrm{DINOv2}}(I) \in \mathbb{R}^{h \times w \times d},$$5 and $$f_{\mathrm{DINOv2}}(I) \in \mathbb{R}^{h \times w \times d},$$6 for MA-ELoFTR, compared with $$f_{\mathrm{DINOv2}}(I) \in \mathbb{R}^{h \times w \times d},$$7 and $$f_{\mathrm{DINOv2}}(I) \in \mathbb{R}^{h \times w \times d},$$8 for XoFTR, and $$f_{\mathrm{DINOv2}}(I) \in \mathbb{R}^{h \times w \times d},$$9 and $p \in \Omega_s$0 for RoMa (Corley et al., 11 Apr 2026). The paper characterizes XoFTR and RoMa as tied at $p \in \Omega_s$1 px mean error, the lowest reported in that table excerpt, while MatchAnything-ELoFTR is described as a close third at $p \in \Omega_s$2 px despite being zero-shot.

The broader benchmark conclusion is that asymmetric transfer is observed: matchers with explicit cross-modal training do not uniformly outperform those without it (Corley et al., 11 Apr 2026). MatchAnything-ELoFTR is central to that conclusion because it is trained on synthetic cross-modal pairs yet only slightly outperforms plain RoMa on some comparisons, while RoMa itself is reported to achieve the same lowest mean error as XoFTR without cross-modal training. The paper explicitly frames the interpretation as a working hypothesis: foundation-model features, specifically DINOv2 backbones, may contribute to modality invariance that partially substitutes for explicit cross-modal supervision (Corley et al., 11 Apr 2026).

This interpretation should not be overstated. The reported evidence is comparative and zero-shot rather than causal. What is established empirically is narrower: MatchAnything-ELoFTR is strong, but its gains over other DINOv2-backed matchers are modest in the tested regime.

5. Protocol sensitivity and deployment behavior

A major result of the benchmark is that deployment protocol choices can dominate matcher-to-matcher differences. Across the evaluated sweep, protocol sensitivity is large enough that accuracy can shift by up to $p \in \Omega_s$3 for a single matcher, and affine geometry alone reduces mean error from $p \in \Omega_s$4 to $p \in \Omega_s$5 px in one cited comparison (Corley et al., 11 Apr 2026). This finding is directly relevant to MatchAnything-ELoFTR because the model’s reported $p \in \Omega_s$6 px performance is inseparable from the tuned inference protocol used to obtain it.

The protocol-sensitivity study comprises 64 runs per matcher. The paper reports that switching from homography to affine lowers mean error by approximately $p \in \Omega_s$7 px on average. Tile size matters more than overlap, with the best results occurring near the matcher’s native training resolution, for example $p \in \Omega_s$8–$p \in \Omega_s$9 px. RANSAC-threshold sweeps from $q \in \Omega_o$0 to $q \in \Omega_o$1 px show non-monotonic trade-offs: strong dense matchers such as MA-ELoFTR and MINIMA-RoMa tolerate tighter thresholds, whereas sparser matchers such as LoFTR and XFeat require looser gating to recover enough inliers (Corley et al., 11 Apr 2026).

The paper also reports normalization ablations. MA-ELoFTR and RoMa are comparatively robust to normalization choice, whereas LoFTR and XFeat vary by up to $q \in \Omega_o$2 in mean error (Corley et al., 11 Apr 2026). This makes MA-ELoFTR comparatively forgiving in deployment, though still not invariant to protocol decisions. The recommended baseline for zero-shot optical–SAR registration in the wild is to preprocess with percentile clipping or CLAHE, resize the long side to at most $q \in \Omega_o$3 px, use $q \in \Omega_o$4 px tiles with at least $q \in \Omega_o$5 px overlap, and estimate geometry with affine RANSAC using reprojection threshold at most $q \in \Omega_o$6 px and minimum inliers $q \in \Omega_o$7 (Corley et al., 11 Apr 2026). Within that baseline, the paper recommends RoMa or MINIMA-RoMa as default and identifies MA-ELoFTR as a close alternative if it is already installed.

A common misconception in matcher benchmarking is that the network alone determines registration quality. The results reported here argue against that view: for MA-ELoFTR, as for other evaluated matchers, geometric model selection, tile resolution, overlap, normalization, and inlier gating are first-order variables rather than ancillary implementation details (Corley et al., 11 Apr 2026).

6. Relation to adjacent cross-modal matching work

Within the satellite-registration study, MatchAnything-ELoFTR functions as an example of a detector-free matcher whose pretraining explicitly targets cross-modal transfer but whose final performance still depends heavily on geometric post-processing and inference protocol (Corley et al., 11 Apr 2026). This places it in a broader class of correspondence architectures that operate through dense feature extraction, coarse cross-attention or assignment, and fine-stage refinement rather than through classical detect-then-describe pipelines.

A related extension of the detector-free paradigm appears in "Single-Frame Point-Pixel Registration via Supervised Cross-Modal Feature Matching" (Han et al., 28 Jun 2025). That work projects a single LiDAR sweep into a 2D intensity image and processes the LiDAR image and camera image with dual non-weight-sharing ResNet-style backbones, followed by a LoFTR-style coarse-level transformer with alternating self-attention and cross-attention layers. It then introduces a repeatability scoring mechanism as a soft visibility prior and reports state-of-the-art results on KITTI, nuScenes, and MIAS-LCEC-TF70 benchmarks (Han et al., 28 Jun 2025). The paper explicitly states that the detector-free pipeline is fully compatible with MatchAnything’s universal matching paradigm and suggests three concrete extensions: precomputing the LiDAR intensity image as a pseudo-modal input to ELoFTR, integrating the repeatability MLP as a visibility-prior head in ELoFTR’s matching module, and leveraging MatchAnything’s large-scale pretraining on heterogeneous modalities before fine-tuning with repeatability supervision (Han et al., 28 Jun 2025).

This connection is informative for interpreting MA-ELoFTR. It suggests that the MatchAnything formulation is not limited to SAR–optical matching as such, but can be understood as part of a more general program of cross-modal dense correspondence learning. The available evidence does not establish that a single universal recipe will solve all modality pairs. It does, however, indicate that detector-free matching with shared or aligned feature spaces, coarse-to-fine refinement, and synthetic or projected pseudo-modal supervision is emerging as a common design pattern across remote sensing and robotic perception.

7. Significance, limitations, and interpretation

MatchAnything-ELoFTR is significant in the benchmark not because it wins unambiguously, but because it sharpens several empirical conclusions. First, it shows that synthetic cross-modal pretraining can produce a strong off-the-shelf matcher for SAR–optical registration, yielding $q \in \Omega_o$8 px mean error and $q \in \Omega_o$9 failure on the labeled SpaceNet9 training scenes under the chosen protocol (Corley et al., 11 Apr 2026). Second, because RoMa attains the same lowest reported mean error as XoFTR without cross-modal training, and because MA-ELoFTR only slightly outperforms plain RoMa in the reported comparisons, the results undermine the simple assumption that explicit cross-modal supervision is always decisive (Corley et al., 11 Apr 2026). The paper’s own wording is appropriately cautious: foundation-model features may contribute to modality invariance that partially substitutes for explicit cross-modal supervision.

The study also clarifies what MatchAnything-ELoFTR is not. It is not presented as a newly introduced architecture in the satellite-registration paper. It is not evaluated with fine-tuning or domain adaptation on the target SAR benchmarks. It is not sufficient, on its own, to eliminate protocol sensitivity. And it is not evidence that all nontraditional matchers transfer well: the same paper reports that 3D-reconstruction matchers such as MASt3R and DUSt3R are highly protocol-sensitive and remain fragile under default settings, with poor performance on orthorectified imagery that has no depth relief (Corley et al., 11 Apr 2026).

From an encyclopedic perspective, MatchAnything-ELoFTR is best understood as a pretrained ELoFTR variant whose importance lies in comparative transfer analysis. It occupies a technically revealing position between explicitly cross-modal training and foundation-feature-based generalization. The benchmark evidence suggests that much of its practical value in SAR–optical registration derives from the combination of a DINOv2-backed detector-free matcher and a carefully tuned affine-RANSAC tiling pipeline, rather than from synthetic cross-modal pretraining alone (Corley et al., 11 Apr 2026).