Are Pretrained Image Matchers Good Enough for SAR-Optical Satellite Registration?

Conference: CVPR 2026 arXiv: 2604.10217 Code: None Area: Remote Sensing Imagery Keywords: SAR-optical registration, image matching, cross-modal, zero-shot transfer, satellite imagery

TL;DR

This paper evaluates 24 families of pretrained image matchers on SAR-optical satellite registration under a zero-shot setting, finding that deployment protocol choices (geometric model, tile size, etc.) can affect accuracy by up to 33×, sometimes surpassing the effect of switching the matcher itself.

Background & Motivation

Background: Cloud cover during disaster response frequently renders optical imagery unavailable, necessitating the registration of SAR images to optical basemaps for generating georeferenced damage assessments. However, state-of-the-art image matchers are designed for indoor, urban, or natural scene imagery.

Limitations of Prior Work: Optical and SAR sensors observe the same scene through fundamentally different physical mechanisms — optical sensors capture reflected light (texture-rich), while SAR captures radar backscatter (speckle noise, layover, radiometric inversion). Whether pretrained matchers can function under such extreme domain shift remains unclear.

Key Challenge: Cross-modal matching requires modality-invariant feature representations, yet pretrained data contains virtually no satellite or SAR imagery.

Goal: Evaluate the cross-modal satellite registration performance of 24 matcher families in a purely zero-shot setting, without any fine-tuning or domain adaptation.

Key Insight: A unified, deterministic evaluation protocol encompassing large-image tile-based inference, robust geometric filtering, and tie-point-anchored metrics.

Core Idea: Cross-modal transfer is asymmetric — explicit cross-modal training does not consistently outperform training on natural images alone, and foundation model features may partially substitute for cross-modal supervision.

Method

Overall Architecture

24 matcher families × unified evaluation protocol (tile-based inference + RANSAC geometric filtering + tie-point metrics) × SpaceNet9 and two additional cross-modal benchmarks. Deployment protocol choices (geometric model, tile size, inlier gating) are systematically ablated.

Key Designs

1. Unified Evaluation Protocol:

   • Function: Ensures all 24 matchers are evaluated under fully comparable conditions.
   • Mechanism: Tile-based inference for handling the very high resolution of satellite imagery, affine RANSAC or fundamental matrix RANSAC for geometric filtering, and tie-point-anchored reprojection error metrics.
   • Design Motivation: Different matcher papers employ different evaluation setups, making direct comparison meaningless without standardization.

2. Deployment Protocol Sensitivity Analysis:

   • Function: Quantifies the impact of hyperparameter choices on registration accuracy.
   • Mechanism: Systematic sweep over protocol parameters including geometric model (affine/fundamental matrix/homography), tile size, and inlier gating threshold. Using affine geometry alone reduces mean error from 12.34 px to 9.74 px. A single matcher can exhibit up to 33× accuracy variation across different protocols.
   • Design Motivation: In practical deployment, protocol selection may matter more than matcher selection.

3. Cross-Modal Transfer Asymmetry Finding:

   • Function: Reveals that explicit cross-modal training is not a prerequisite for strong performance.
   • Mechanism: XoFTR (trained on visible-thermal pairs) and RoMa (no cross-modal training) both achieve the lowest mean error of 3.0 px. MatchAnything-ELoFTR (trained on synthetic cross-modal pairs) follows closely at 3.4 px. DINOv2 foundation model features may provide partial modality invariance.
   • Design Motivation: Challenges the intuition that cross-modal tasks necessarily require cross-modal training.

Loss & Training

This is a purely zero-shot evaluation study; no training is involved. All matchers use their official pretrained weights.

Key Experimental Results

Main Results

Matcher	SpaceNet9 Mean Error (px)	Cross-Modal Training
XoFTR	3.0	Yes (visible–thermal)
RoMa	3.0	No
MatchAnything-ELoFTR	3.4	Yes (synthetic cross-modal)
MASt3R/DUSt3R	Protocol-sensitive	No (3D reconstruction)

Ablation Study

Protocol Choice	Mean Error Change	Note
Affine geometry vs. others	12.34→9.74 px	21% reduction
Tile size variation	Up to 33× difference	For individual matchers
Inlier gating variation	Significant impact	Both overly strict and overly loose settings degrade performance

Key Findings

• The impact of deployment protocol choices can exceed that of switching matchers — affine geometry alone reduces error by 21%.
• 3D reconstruction-based matchers (MASt3R/DUSt3R) are highly fragile under default settings and are heavily dependent on protocol configuration.
• DINOv2 foundation model features may provide a form of implicit modality invariance.

Highlights & Insights

• "Protocol over algorithm" finding: For practitioners, optimizing the deployment protocol may be more effective than replacing the matcher.
• Cross-modal transfer asymmetry: Counterintuitively, RoMa achieves the lowest error without any cross-modal training.
• Hypothesis of implicit modality invariance in DINOv2: Foundation model features may inherently generalize across modalities.

Limitations & Future Work

• Only zero-shot performance is evaluated; the effect of few-shot fine-tuning remains unexplored.
• Scene coverage in SpaceNet9 may be limited (primarily urban areas).
• The hypothesis of implicit modality invariance in DINOv2 requires deeper mechanistic analysis.

Related Work & Insights

• vs. RemoteCLIP: RemoteCLIP achieves domain adaptation through large-scale remote sensing pretraining, whereas this paper demonstrates that zero-shot transfer may already be sufficient.
• vs. LoFTR/ELoFTR: Standard natural-image matchers exhibit inconsistent performance on cross-modal satellite registration; deployment protocol is the critical factor.

Rating

• Novelty: ⭐⭐⭐ Empirical study, but findings are valuable.
• Experimental Thoroughness: ⭐⭐⭐⭐⭐ 24 matcher families × multiple protocols × multiple benchmarks — extremely comprehensive.
• Writing Quality: ⭐⭐⭐⭐ In-depth analysis with a practical orientation.
• Value: ⭐⭐⭐⭐ Provides direct guidance for real-world deployment in applications such as disaster response.

Related Papers

• [CVPR 2026] SDF-Net: Structure-Aware Disentangled Feature Learning for Optical-SAR Ship Re-identification
• [ICCV 2025] WildSAT: Learning Satellite Image Representations from Wildlife Observations
• [ICLR 2026] TAMMs: Change Understanding and Forecasting in Satellite Image Time Series with Temporal-Aware Multimodal Models
• [CVPR 2026] Joint and Streamwise Distributed MIMO Satellite Communications with Multi-Antenna Ground Users
• [AAAI 2026] Perceive, Act and Correct: Confidence Is Not Enough for Hyperspectral Classification