acceptodds
Under review as a conference paper at ICLR 2027

GeoCFR: Multimodal Remote Sensing Image Registration via Structural Representation Learning and Posterior Aggregation

Abstract

High-precision registration of multimodal remote sensing images (MRSIs) is fundamental to multi-source information fusion and Earth observation applications. However, cross-modal appearance differences combined with large rotations, translations, and scale changes make reliable correspondences difficult to establish. Sparse methods can handle large initial displacements, but cross-modal differences can result in too few reliable correspondences or uneven spatial coverage; dense methods provide more candidates but do not fully exploit candidate distributions in local responses for accurate global geometry estimation. We therefore propose GeoCFR, which combines cross-modal structural representation learning with local matching posterior aggregation to progressively refine a global affine transformation. The regional moment–relational spectrum descriptor (RMSD) uses regional statistics and multi-ring relationships to obtain an affine initialization; the Posterior-Aggregating Tokens for Affine Refinement (PATAR) module aggregates local candidate distributions and geometric information to predict global affine residuals at two feature scales. We also construct WH-M3R, a dataset covering Wuhan with 343,758 valid cross-modal image pairs across 13 data layers at spatial resolutions of 0.5, 2, 5, and 10 m. Match-30K is used for training and validation, and Match-1K for testing. Under the High protocol on WH-M3R-Match-1K, GeoCFR improves the correct matching rate (CMR) at the 1-, 2-, and 3-pixel thresholds by 7.7, 6.9, and 5.8 percentage points over the strongest baseline at each threshold. Under the High protocol, GeoCFR achieves leading overall zero-shot performance on SOMA-1M-Test, RoadScene, and OSdataset, demonstrating transfer across datasets under large geometric changes.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
Image Matching, Visual Correspondence, Multimodal Remote Sensing Image Registration, Multiresolution Dataset

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 3 reviews, above 27% of the others.
Reviewer_mTQsOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes GeoCFR for multimodal remote-sensing image registration under large geometric changes. It first uses RMSD to learn cross-modal structural features and estimate an initial affine transformation, then uses PATAR to aggregate local matching posteriors and their geometric information to progressively refine the global transformation. The authors also construct WH-M3R, a dataset spanning multiple remote-sensing data layers and spatial resolutions. Experiments show that GeoCFR outperforms the evaluated baselines under three geometric-transformation protocols on WH-M3R and achieves promising zero-shot transfer performance on SOMA-1M-Test, RoadScene, and OSdataset.

Critical strengths
  1. The method is conceptually well motivated and addresses key challenges in registration.
  2. The authors compare handcrafted, sparse, semi-dense, and dense methods under multiple geometric-transformation protocols and conduct zero-shot evaluation on external datasets, demonstrating strong registration performance.
Critical weaknesses
  1. The current experiments do not sufficiently validate registration under general affine transformations.
  2. The authors provide a useful manual assessment of WH-M3R, but the results indicate residual misalignment in the original image pairs and substantial variation in alignment quality across data layers.
Questions for authors
  1. Can the authors provide experiments with general affine transformations?
  2. How much do residual alignment errors and differences in data-layer difficulty affect the conclusions?
  3. How can the gains from PATAR be distinguished from those due to additional model capacity?
4 comments
AuthorsOpenReview ↗

Overall response

Thank you for recognizing our motivation and raising questions about general affine evaluation, data-layer differences, and model capacity. The added affine tests show direct transfer of the existing models to broader geometric conditions. Data-layer exclusion and label-perturbation analyses support the comparative conclusions under changes in data composition and small reference shifts. We also completed a statistics-only comparison with the same architecture and parameter count. The full posterior achieves higher CMR on all nine protocol–threshold combinations.

AuthorsOpenReview ↗

W1 / Q1: Evaluation under general affine transformations

We added RoMa following Reviewer Sja7's suggestion. GeoCFR remains best on all nine CMR metrics under the original Low, Medium, and High protocols.

For general affine transformations, we added tests without further training or fine-tuning. GeoCFR ranks first on two metrics and second on seven among the five compared methods. We use the same 1,000 source image pairs and retain the existing weights for every method. The three protocols add shear of ±3°/±6°/±9° and area-preserving anisotropic scaling, with maximum axis ratios of 1.03/1.06/1.09, respectively.

Table M1: Additional general affine tests (CMR, %).

LowMediumHighMethod1p˜x2p˜x3p˜x1p˜x2p˜x3p˜x1p˜x2p˜x3p˜xRAFT1.7/2.618.4/20.748.2/51.20.2/0.54.2/5.615.8/20.10.0/0.30.9/3.84.9/11.7GDROS2.720.045.90.86.418.60.02.59.0RoMa3.0/11.9‾35.5/\mathbfsf45.956.6/\mathbfsf62.91.8/5.8‾20.1/\mathbfsf30.540.6/\mathbfsf47.71.9/\mathbfsf6.113.2/\mathbfsf21.525.6/\mathbfsf36.6WAFT1.4/2.011.9/12.928.2/30.20.3/1.33.8/4.58.4/9.20.0/0.21.4/2.24.4/4.7GeoCFR\mathbfsf15.134.3‾52.1‾\mathbfsf9.521.2‾32.1‾2.2‾8.1‾16.0‾{\small\def\R#1{\mathsf{#1}}\def\B#1{\mathbfsf{#1}}\begin{array}{lccccccccc}\hline&&\mathclap{\begin{array}{c}\R{Low}\\\\[-.65em]\rule{14.4em}{.4pt}\end{array}}&&&\mathclap{\begin{array}{c}\R{Medium}\\\\[-.65em]\rule{13.9em}{.4pt}\end{array}}&&&\mathclap{\begin{array}{c}\R{High}\\\\[-.65em]\rule{13.9em}{.4pt}\end{array}}&\\\\\R{Method}&\R{1\~px}&\R{2\~px}&\R{3\~px}&\R{1\~px}&\R{2\~px}&\R{3\~px}&\R{1\~px}&\R{2\~px}&\R{3\~px}\\\\\hline\R{RAFT}&\sf1.7/2.6&\sf18.4/20.7&\sf48.2/51.2&\sf0.2/0.5&\sf4.2/5.6&\sf15.8/20.1&\sf0.0/0.3&\sf0.9/3.8&\sf4.9/11.7\\\\\R{GDROS}&\sf2.7&\sf20.0&\sf45.9&\sf0.8&\sf6.4&\sf18.6&\sf0.0&\sf2.5&\sf9.0\\\\\R{RoMa}&\sf3.0/\underline{11.9}&\sf35.5/\B{45.9}&\sf56.6/\B{62.9}&\sf1.8/\underline{5.8}&\sf20.1/\B{30.5}&\sf40.6/\B{47.7}&\sf1.9/\B{6.1}&\sf13.2/\B{21.5}&\sf25.6/\B{36.6}\\\\\R{WAFT}&\sf1.4/2.0&\sf11.9/12.9&\sf28.2/30.2&\sf0.3/1.3&\sf3.8/4.5&\sf8.4/9.2&\sf0.0/0.2&\sf1.4/2.2&\sf4.4/4.7\\\\\R{GeoCFR}&\B{15.1}&\sf\underline{34.3}&\sf\underline{52.1}&\B{9.5}&\sf\underline{21.2}&\sf\underline{32.1}&\sf\underline{2.2}&\sf\underline{8.1}&\sf\underline{16.0}\\\\\hline\end{array}}

\stylefont−size:12pxNote: Original and affine-fitted outputs appear before and after the slash. Rankings use each method’s better output. Bold and\stylefont−size:12pxunderlining mark the best and second-best methods, respectively.{\def\N#1{\style{font-size:12px}{\textsf{#1}}}\begin{array}{l}\N{Note: Original and affine-fitted outputs appear before and after the slash. Rankings use each method's better output. Bold and}\\\\\N{underlining mark the best and second-best methods, respectively.}\end{array}}

Rankings use the better of each method's original and Affine outputs. Under Low and Medium, CMR@1 reaches 15.1% and 9.5%, exceeding RoMa+Affine by 3.2 and 3.7 percentage points, respectively. GeoCFR scores higher than RAFT, GDROS, WAFT, and their corresponding outputs on all nine metrics. RoMa+Affine achieves higher values on the remaining seven metrics. RoMa uses the larger DINOv2 ViT-L/14 pretrained backbone; its architecture and pretraining background differ from GeoCFR's. In our timing test, GeoCFR takes 107.8 ms per pair, compared with 211.7 ms for RoMa (Appendix G, Table 17).

The paired experiment under Low separates the effects of shear and anisotropic scaling. CMR@3 is 77.7% with anisotropic scaling alone and 58.1% with shear alone, versus the 80.5% similarity-transform reference. Shear changes angular and radial relationships in MRRS ring sampling, which may make coarse matching harder under unseen deformations. PATAR's six-parameter updates retain the ability to represent these deformations. Together with the Low/Medium CMR@1 advantages, the added tests support the full framework's direct transfer to unseen general affine transformations.

Revised sections: p. 7, Table 4; p. 26, Appendix H (General-Affine Generalization), Tables 18–19; p. 24, Appendix G, Table 17.

AuthorsOpenReview ↗

W2 / Q2: Effects of residual misalignment and data-layer differences on the conclusions

We separately tested how data-layer composition and evaluation-label shifts affect comparisons between methods. Table M2 summarizes the key High-protocol results.

Data-layer exclusion. The original test set contains 10 target data layers, with 100 pairs per layer and 1,000 pairs in total. Depth is generated from the reference RGB images. We therefore first excluded Depth, leaving 900 pairs, and then excluded Depth, Map, and DEM together, leaving 700 pairs. This tests whether the overall advantage depends on these derived or difficult layers. We recomputed the metrics for 17 outputs from 12 methods on the same retained pairs, including original and Affine outputs from RoMa and MINIMA-RoMa. After excluding Depth, GeoCFR remains ahead on all nine CMR metrics across the three protocols. Under High, excluding Depth, Map, and DEM gives CMR@1/2/3 of 34.43/63.86/76.57%. The margins over the strongest baseline are 9.43/8.43/6.29 percentage points, and all corresponding intervals are above zero.

Evaluation-label translations. Manual point discrepancies combine residual misalignment and cross-modal localization uncertainty. We fixed each method's predictions on 1,000 pairs and translated only the correspondence positions used for scoring. This shows how CMR changes and whether method rankings at strict pixel thresholds change. Table M2 reports the 0.5 px and 1 px results.

Table M2: Data-layer exclusion and evaluation-label translation under High. CMR@1 is the proportion of pairs whose mean EPE is at most 1 px. Δ is GeoCFR's CMR@1 minus the highest baseline CMR@1 for each setting, in percentage points. Brackets show 95% confidence intervals.

{\small\def\R#1{\mathsf{#1}}\begin{array}{lcc}\hline\begin{array}{l}\R{Analysis\~setting}\\\\\R{(sample\~size)}\end{array}&\R{GeoCFR\~CMR@1\~(\\%)}&\R{ΔCMR@1\~(pp)\~[95\\%\~CI]}\\\\\hline\begin{array}{l}\R{Excluding\~Depth\~(900}\\\\\R{pairs)}\end{array}&\sf27.11&\sf+7.33\\;[4.89,9.78]\\\\\begin{array}{l}\R{Excluding\~Depth,\~Map,\~and}\\\\\R{DEM\~(700\~pairs)}\end{array}&\sf34.43&\sf+9.43\\;[6.29,12.57]\\\\\begin{array}{l}\R{Label\~translation\~0.5\~px}\\\\\R{(1000\~pairs)}\end{array}&\sf29.89&\sf+7.48\\;[5.86,8.99]\\\\\begin{array}{l}\R{Label\~translation\~1\~px}\\\\\R{(1000\~pairs)}\end{array}&\sf15.12&\sf+3.49\\;[2.63,4.35]\\\\\hline\end{array}}

Data-layer exclusion uses the original evaluation region and 10,000 paired bootstrap resamples stratified by data layer. Label translation uses a fixed common valid region and 20 shared direction realizations. It compares RAFT, GDROS, WAFT, RoMa, and their available Affine outputs, using 2,000 paired bootstrap resamples.

Both analyses resample image pairs and reselect the strongest baseline in every resample. The intervals are pointwise 95% intervals. At label translations of 0, 0.25, 0.5, and 1 px, the High CMR@1/2/3 gaps and their intervals remain above zero.

Coarse-to-fine analysis on Map and DEM. Each layer has 100 pairs in the original test set. Under High, PATAR refinement raises Map/DEM CMR@3 from 32%/8% to 39%/13%, recovering 10/5 previously unsuccessful pairs, respectively. Three Map pairs change from success to failure; no DEM pairs do. For Map and DEM, 58 and 87 pairs, respectively, remain above the 3 px threshold at both stages. The exclusion and label-translation analyses show that GeoCFR retains its comparative advantage after excluding the relevant layers or applying the tested small label translations. The translation experiment does not estimate the actual residual-error distribution.

Revised sections: pp. 16–17, Appendix A.2.1 (Sensitivity to Evaluation-Label Perturbations), Table 10; pp. 22–24, Appendix E (Results by Data-Layer Pairing), Tables 14–16.

AuthorsOpenReview ↗

Q3: Posterior-information gains versus model capacity in PATAR

We completed a comparison between statistics-only encoding and the full posterior using the same architecture and parameter count. The full posterior gives higher CMR on all nine metrics across the three protocols. Both variants retain the same evidence encoder, tokens, attention layers, residual head, and P4→P2 pipeline. Each has 18.35M total parameters and 1.64M trainable parameters, including 689,820 in the refinement branch.

The statistics-only variant replaces candidate probabilities with a discrete encoding derived only from the original mean, variance trace, and valid-candidate mask. The full-posterior variant retains the original candidate probabilities. Both variants keep the same coordinate, original mean, and variance channels. This compares the information in the full candidate distribution with that retained by low-order statistics.

Table M3: Statistics-only encoding versus the full posterior with the same architecture and parameter count (CMR@1/2/3, %).

MethodLowMediumHighStatistics-onlye˜ncoding45.3/68.7/78.236.9/61.2/72.133.3/56.3/67.2Fullp˜osterior\mathbfsf48.5/\mathbfsf71.0/\mathbfsf80.7\mathbfsf40.2/\mathbfsf63.5/\mathbfsf74.1\mathbfsf34.3/\mathbfsf57.0/\mathbfsf68.7{\small\def\R#1{\mathsf{#1}}\def\B#1{\mathbfsf{#1}}\begin{array}{lccc}\hline\R{Method}&\R{Low}&\R{Medium}&\R{High}\\\\\hline\R{Statistics\text{-}only\~encoding}&\sf45.3/68.7/78.2&\sf36.9/61.2/72.1&\sf33.3/56.3/67.2\\\\\R{Full\~posterior}&\sf\B{48.5}/\B{71.0}/\B{80.7}&\sf\B{40.2}/\B{63.5}/\B{74.1}&\sf\B{34.3}/\B{57.0}/\B{68.7}\\\\\hline\end{array}}

The full posterior improves CMR@1 by 3.2/3.3/1.0 percentage points under Low/Medium/High. Its CMR@3 gains are 2.5/2.0/1.5 percentage points, respectively. Mean image-pair EPE is also lower under all three protocols.

Both variants use the same PATAR architecture and 689,820 parameters in the refinement branch. This removes the parameter-count difference introduced by Peak-Affine's different estimator in the original Table 6. At the same model capacity, the full posterior gives higher CMR than statistics-only encoding on all nine metrics. CMR@1 gains reach 3.3 percentage points.

Revised sections: p. 9, Section 5.3 (Ablation Studies), Table 6 and the “Posterior aggregation” paragraph.

Reviewer_TwiPOpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes GeoCFR, a coarse-to-fine framework for multimodal remote sensing image registration under large geometric transformations. The method introduces a regional moment–relational spectrum descriptor (RMSD) to learn cross-modal structural representations for keypoint matching and initial affine estimation. It further develops Posterior-Aggregating Tokens for Affine Refinement (PATAR), which aggregates local matching probability distributions and geometric information to progressively refine a global affine transformation at two feature scales. The paper also introduces WH-M3R, a multimodal remote sensing dataset containing 343,758 image pairs across 13 data layers and four spatial resolutions. Experiments are conducted on WH-M3R under three levels of geometric transformations, with additional zero-shot evaluations on SOMA-1M-Test, RoadScene, and OSdataset. The results demonstrate improvements over several sparse, semi-dense, and dense matching baselines, particularly at strict registration accuracy thresholds. Ablation studies further support the contributions of the structural descriptor and posterior-based refinement.

Critical strengths

1.Effective integration of structural representation learning and posterior-based geometric refinement. The combination of RMSD and PATAR provides a meaningful approach to cross-modal registration. In particular, aggregating full local matching distributions rather than relying on individual correlation peaks is well motivated and addresses an important source of ambiguity in multimodal matching. The ablation results demonstrate substantial improvements from posterior aggregation and successive refinement, supporting the effectiveness of the proposed design. 2.Extensive experimental evaluation and a potentially valuable multimodal dataset. WH-M3R covers diverse remote sensing data layers and spatial resolutions, providing a useful resource for studying cross-modal registration. The paper evaluates GeoCFR against multiple categories of matching methods under increasingly challenging transformations and includes zero-shot evaluations on three external datasets. The consistent improvements, especially under the High protocol, provide encouraging evidence of the method's effectiveness and generalization capability.

Critical weaknesses
  1. The geometric evaluation does not fully validate the claimed affine registration capability. Although GeoCFR predicts a general six-parameter affine transformation, the Low, Medium, and High protocols only consider rotation, translation, and isotropic scaling, explicitly excluding shear and anisotropic scaling. Consequently, the experiments primarily validate robustness to similarity transformations rather than general affine transformations. This discrepancy limits the evidence supporting the proposed affine refinement framework, particularly for practical registration scenarios involving more complex geometric distortions.
  2. The reported overall performance masks substantial differences across modalities, while dataset alignment uncertainty complicates the interpretation of pixel-level accuracy. Appendix Table 11 shows that performance varies significantly across data layers. Under the High protocol, CMR@1 reaches 99% for Depth but only 1% for DEM and 2% for Map. Since Depth is derived from the reference RGB imagery, its near-perfect performance may not adequately reflect registration across independent sensing modalities. Furthermore, the manual alignment assessment reports a median pair-level error of 1.76 pixels, raising concerns about the reliability of evaluating registration at a strict 1-pixel threshold. Although the authors acknowledge these limitations and provide supplementary consistency checks, their implications for the main conclusions require further investigation.
Questions for authors
  1. Can the authors provide additional experiments involving shear and anisotropic scaling to evaluate the full six-parameter affine estimation capability? Since the current transformation protocols are restricted to similarity transformations, such experiments would clarify whether PATAR generalizes to more complex affine distortions.
  2. Can the authors report aggregate results excluding the Depth modality and provide a more detailed analysis of the challenging Map and DEM cases? This would help determine whether the reported improvements remain consistent across genuinely heterogeneous sensing modalities.
  3. How sensitive are the reported CMR@1 results to residual alignment errors in WH-M3R? Given the uncertainty identified in the manual correspondence assessment, additional analysis of evaluation-label accuracy and its effect on strict pixel-level metrics would strengthen confidence in the conclusions.
4 comments
AuthorsOpenReview ↗

Overall response

Thank you for your detailed comments on geometric validation and evaluation reliability. We added general affine tests using the existing models, without further training. Taking each method's better output, GeoCFR ranks first in two settings and second in seven. After excluding Depth, it still leads on all nine CMR settings across the three protocols. Under High, its margins remain positive after jointly excluding Depth, Map, and DEM, with all three confidence intervals above zero. We also shifted the evaluation labels by 0.25, 0.5, or 1 px under High. GeoCFR retains its CMR@1 lead at each magnitude, and the corresponding 95% confidence intervals remain above zero.

AuthorsOpenReview ↗

W1 / Q1: Direct generalization to general affine transformations

Our original training and main evaluation protocols drew on the rotation, translation, and scaling settings used by GDROS and CRFT. These settings appear in GDROS (TGRS 2025, Section IV-A) and CRFT (CVPR 2026, Section 4.1). We use rotation, translation, and isotropic scaling to provide controlled Low, Medium, and High difficulty levels for all baselines. PATAR retains the full affine representation through six Lie algebra residual parameters and the matrix exponential update (Eq. 7). This representation includes shear and anisotropic scaling.

We also added RoMa at Reviewer Sja7's suggestion. With RoMa included, GeoCFR still achieves the highest CMR in all nine settings under the original Low, Medium, and High protocols.

We added separate and combined tests for shear and anisotropic scaling, together with a paired analysis of the two components. All methods use their existing weights, without further training or fine-tuning for these tests.

We generated three general affine test sets from the same 1,000 source pairs, retaining the original parameter ranges and overlap constraints. The added shear ranges are ±3°/±6°/±9° under Low/Medium/High, respectively. We also add area-preserving anisotropic scaling, with maximum axis ratios of 1.03/1.06/1.09, respectively.

Table T1: Additional general affine tests (CMR, %). We compare strong representative methods from the original protocol and the added RoMa, covering five methods and eight outputs.

LowMediumHighMethod1p˜x2p˜x3p˜x1p˜x2p˜x3p˜x1p˜x2p˜x3p˜xRAFT1.7/2.618.4/20.748.2/51.20.2/0.54.2/5.615.8/20.10.0/0.30.9/3.84.9/11.7GDROS2.720.045.90.86.418.60.02.59.0RoMa3.0/11.9‾35.5/\mathbfsf45.956.6/\mathbfsf62.91.8/5.8‾20.1/\mathbfsf30.540.6/\mathbfsf47.71.9/\mathbfsf6.113.2/\mathbfsf21.525.6/\mathbfsf36.6WAFT1.4/2.011.9/12.928.2/30.20.3/1.33.8/4.58.4/9.20.0/0.21.4/2.24.4/4.7GeoCFR\mathbfsf15.134.3‾52.1‾\mathbfsf9.521.2‾32.1‾2.2‾8.1‾16.0‾{\small\def\R#1{\mathsf{#1}}\def\B#1{\mathbfsf{#1}}\begin{array}{lccccccccc}\hline&&\mathclap{\begin{array}{c}\R{Low}\\\\[-.65em]\rule{14.4em}{.4pt}\end{array}}&&&\mathclap{\begin{array}{c}\R{Medium}\\\\[-.65em]\rule{13.9em}{.4pt}\end{array}}&&&\mathclap{\begin{array}{c}\R{High}\\\\[-.65em]\rule{13.9em}{.4pt}\end{array}}&\\\\\R{Method}&\R{1\~px}&\R{2\~px}&\R{3\~px}&\R{1\~px}&\R{2\~px}&\R{3\~px}&\R{1\~px}&\R{2\~px}&\R{3\~px}\\\\\hline\R{RAFT}&\sf1.7/2.6&\sf18.4/20.7&\sf48.2/51.2&\sf0.2/0.5&\sf4.2/5.6&\sf15.8/20.1&\sf0.0/0.3&\sf0.9/3.8&\sf4.9/11.7\\\\\R{GDROS}&\sf2.7&\sf20.0&\sf45.9&\sf0.8&\sf6.4&\sf18.6&\sf0.0&\sf2.5&\sf9.0\\\\\R{RoMa}&\sf3.0/\underline{11.9}&\sf35.5/\B{45.9}&\sf56.6/\B{62.9}&\sf1.8/\underline{5.8}&\sf20.1/\B{30.5}&\sf40.6/\B{47.7}&\sf1.9/\B{6.1}&\sf13.2/\B{21.5}&\sf25.6/\B{36.6}\\\\\R{WAFT}&\sf1.4/2.0&\sf11.9/12.9&\sf28.2/30.2&\sf0.3/1.3&\sf3.8/4.5&\sf8.4/9.2&\sf0.0/0.2&\sf1.4/2.2&\sf4.4/4.7\\\\\R{GeoCFR}&\B{15.1}&\sf\underline{34.3}&\sf\underline{52.1}&\B{9.5}&\sf\underline{21.2}&\sf\underline{32.1}&\sf\underline{2.2}&\sf\underline{8.1}&\sf\underline{16.0}\\\\\hline\end{array}}

\stylefont−size:12pxNote: Slash-separated entries denote raw/affine-fitted outputs. Rankings use each method’s better output; bold and underlined\stylefont−size:12pxvalues mark the best and second-best methods, respectively. CMR@t is the percentage of image pairs whose mean EPE over valid\stylefont−size:12pxpixels does not exceed t pixels.{\def\N#1{\style{font-size:12px}{\textsf{#1}}}\begin{array}{l}\N{Note: Slash-separated entries denote raw/affine-fitted outputs. Rankings use each method's better output; bold and underlined}\\\\\N{values mark the best and second-best methods, respectively. CMR@t is the percentage of image pairs whose mean EPE over valid}\\\\\N{pixels does not exceed t pixels.}\end{array}}

Taking each method's better raw/Affine output, GeoCFR ranks first in two settings and second in seven. Its strict CMR@1 values under Low and Medium are 15.1% and 9.5%, exceeding RoMa+Affine by 3.2 and 3.7 percentage points, respectively. It exceeds RAFT, GDROS, WAFT, and their available Affine outputs on all nine metrics. RoMa+Affine leads on the other seven metrics. RoMa uses a larger DINOv2 ViT-L/14 pretrained backbone. In our timing test, GeoCFR takes 107.8 ms per pair, compared with 211.7 ms for RoMa (Appendix G, Table 17). Without training for the added transforms, GeoCFR remains in the top two across all nine settings. It also retains the lead at the 1 px threshold under Low and Medium.

Table T2: Paired analysis of the added components (Low, CMR, %). Source pairs, sampled rotation/translation/isotropic scale, and crop context are fixed. All four conditions use a common valid region.

{\small\def\R#1{\mathsf{#1}}\begin{array}{lcc}\hline\R{Condition}&\R{CMR@1\~(\\%)}&\R{CMR@3\~(\\%)}\\\\\hline\begin{array}{l}\R{Similarity}\\\\\R{transformations}\end{array}&\sf47.0&\sf80.5\\\\\R{+\~Shear}&\sf20.1&\sf58.1\\\\\R{+\~Anisotropic\~scaling}&\sf35.5&\sf77.7\\\\\R{+\~Both}&\sf14.8&\sf52.0\\\\\hline\end{array}}

Within the tested anisotropic scaling range, CMR@3 is 77.7%, retaining 96.5% of the similarity-transform result. Shear causes a larger decrease in the tested settings. Shear changes the angular and radial relations used by MRRS ring sampling (Eq. 2). Since the original task training did not include shear, this may make coarse matching more difficult. PATAR's six-parameter updates retain the ability to represent these distortions. In the general affine tests, GeoCFR's Low/Medium CMR@1 remains above RoMa+Affine (15.1% vs. 11.9%; 9.5% vs. 5.8%).

Revised sections: p. 7, Table 4; p. 26, Appendix H (General-Affine Generalization), Tables 18–19; p. 24, Appendix G, Table 17.

AuthorsOpenReview ↗

W2 / Q2: Excluding Depth and analyzing Map/DEM

We address this concern through aggregate comparisons and separate analyses of Map and DEM. Depth is derived from the reference RGB images, so we first exclude this layer to test whether the aggregate advantage depends on it. Under High, we also evaluate two further subsets: one excluding Map and DEM, and one excluding Depth, Map, and DEM. The subset excluding all three tests whether GeoCFR's advantage persists on the remaining SAR, multispectral, hyperspectral, and panchromatic image pairs. We then compare all methods on Map and DEM separately and examine the changes from coarse matching to refinement.

We recomputed the metrics for all 17 method/output variants on the same retained pairs, counting failed estimates in the denominator. After excluding Depth, GeoCFR still leads on all nine CMR settings across the three protocols. Table T3 reports the complete High-protocol comparisons for all four subsets.

Table T3: Data-layer exclusion analysis under High. CMR is in %; Δ is the margin over the strongest included baseline, in percentage points.

{\small\def\R#1{\mathsf{#1}}\begin{array}{lcccc}\hline\R{High\~subset\~(N)}&\R{GeoCFR\~@1/@2/@3\~(\\%)}&\R{Δ@1\~[95\\%\~CI]}&\R{Δ@2\~[95\\%\~CI]}&\R{Δ@3\~[95\\%\~CI]}\\\\\hline\R{All\~(1000)}&\sf34.30/57.00/68.70&\sf+7.70\\;[5.40,10.00]&\sf+6.90\\;[4.60,8.50]&\sf+5.80\\;[3.80,7.80]\\\\\R{Excluding\~Depth\~(900)}&\sf27.11/52.33/65.33&\sf+7.33\\;[4.89,9.78]&\sf+7.00\\;[4.56,9.00]&\sf+6.56\\;[4.33,8.78]\\\\\R{Excluding\~Map/DEM\~(800)}&\sf42.50/68.25/79.38&\sf+9.62\\;[6.75,12.38]&\sf+8.13\\;[5.50,10.00]&\sf+5.38\\;[3.25,7.50]\\\\\begin{array}{l}\R{Excluding\~Depth,\~Map,\~and}\\\\\R{DEM\~(700)}\end{array}&\sf34.43/63.86/76.57&\sf+9.43\\;[6.29,12.57]&\sf+8.43\\;[5.57,10.86]&\sf+6.29\\;[4.00,8.71]\\\\\hline\end{array}}

{\def\N#1{\style{font-size:12px}{\textsf{#1}}}\begin{array}{l}\N{Note: The intervals use 10,000 paired bootstrap resamples of image pairs, stratified by data layer. The strongest baseline is}\\\\\N{reselected in each resample, and all 12 pointwise 95\\% intervals are above zero.}\end{array}}

The 700 pairs remaining after excluding Depth, Map, and DEM cover SAR, multispectral, hyperspectral, and panchromatic data layers. GeoCFR leads by 9.43/8.43/6.29 percentage points at the three thresholds.

To examine the difficult Map and DEM layers, we added CMR@5 while retaining the original 1/2/3 px metrics. We focus on registration performance at 3/5 px. Under High, GeoCFR reaches 39%/64% CMR@3/5 on Map, exceeding the strongest baseline by 5/10 percentage points. On DEM, it reaches 13%/26%, with margins of 5/1 percentage points. GeoCFR achieves the highest value in seven of the eight layer–threshold settings.

Stage comparisons show how refinement helps at the 3 px threshold. On Map, CMR increases from 32% to 39%. Ten pairs change from failure to success, while 3 change from success to failure; 58 fail at both stages. On DEM, it increases from 8% to 13%, recovering 5 pairs with no regression; 87 pairs fail at both stages. Refinement lowers EPE on 60/100 Map pairs and 81/100 DEM pairs. Refinement recovers some failures, but many pairs remain difficult.

The effect of residual alignment uncertainty on strict metrics is further examined through label perturbations in Q3 below.

Revised sections: pp. 22–24, Appendix E (Results by Data-Layer Pairing), Tables 14–16.

AuthorsOpenReview ↗

Q3: Residual alignment uncertainty and strict pixel thresholds

Under High, GeoCFR retains its CMR@1 lead over the strongest baseline after the tested 0.25, 0.5, and 1 px label translations. The added analysis measures how label shifts affect absolute CMR and whether they change the comparison between methods.

The manual assessment in Appendix A.2 reports a median pair-level discrepancy of 1.76 px. This combines residual misalignment with uncertainty in placing landmarks across modalities. GeoCFR, WAFT, and MINIMA-RoMa report median model-assisted consistency residuals of 0.68–0.78 px. MINIMA-RoMa was not trained on WH-M3R, and its median uses successful estimates. These residuals provide further evidence of consistency, but do not directly measure the true label error. The synthetic transform is known, but its interpretation as a physical correspondence still depends on the original pre-alignment quality.

We hold predictions fixed for 1,000 pairs per protocol and focus on spatially constant label translations of 0, 0.25, 0.5, and 1 px. We retain the fixed common valid region and 20 random-direction realizations shared across methods and magnitudes. Under High, baselines include RAFT, GDROS, WAFT, RoMa, and the Affine outputs of RAFT/WAFT/RoMa. We use 2,000 paired bootstrap resamples of image pairs, stratified by data layer, reselecting the strongest baseline each time. The intervals are pointwise 95% intervals conditional on the fixed 20 direction realizations.

Table T4: GeoCFR CMR@1 (%) across three protocols under 0–1 px label translations, and its High margin in percentage points. All values use the common valid region.

{\small\def\R#1{\mathsf{#1}}\begin{array}{lcccc}\hline\R{Shift\~(px)}&\R{Low}&\R{Medium}&\R{High}&\R{High:\~gap\~[95\\%\~CI]}\\\\\hline\sf0&\sf48.60&\sf40.20&\sf34.30&\sf+7.70\\;[5.40,9.90]\\\\\sf0.25&\sf46.70&\sf38.63&\sf32.87&\sf+7.48\\;[5.57,9.36]\\\\\sf0.5&\sf41.92&\sf35.09&\sf29.89&\sf+7.48\\;[5.86,8.99]\\\\\sf1&\sf20.13&\sf17.14&\sf15.12&\sf+3.49\\;[2.63,4.35]\\\\\hline\end{array}}

{\def\N#1{\style{font-size:12px}{\textsf{#1}}}\begin{array}{l}\N{Note: Zero translation is also scored on the common valid region, giving Low CMR@1 of 48.60\\%. It is 48.50\\% on the original}\\\\\N{evaluation region.}\end{array}}

Under High, GeoCFR still leads by 7.48 and 3.49 percentage points at 0.5 and 1 px, respectively. The paired 95% intervals are [5.86, 8.99] and [2.63, 4.35]. High CMR@2/3 margins and their pointwise 95% intervals are also above zero at all four main magnitudes.

In additional 1.5–2 px stress tests, the High CMR@1 gap intervals include zero. The CMR@2/3 margins and their corresponding intervals remain above zero.

These controlled perturbations measure evaluation sensitivity, not the actual distribution of residual misalignment.

External tests provide independent performance evidence. Under High on SOMA-1M-Test, GeoCFR achieves 13.90% CMR@1, above WAFT+Affine at 10.60% and RoMa+Affine at 4.30%. On OSdataset, GeoCFR reaches 12.97%/30.19% CMR@2/3, leading the comparison in Table 5 of the revised manuscript.

Revised sections: pp. 16–17, Appendix A.2.1 (Sensitivity to Evaluation-Label Perturbations), Table 10; p. 8, Section 5.2 (Results), Table 5.

Reviewer_Sja7OpenReview ↗
Rating
2
Confidence
2
Informativeness
1
Presentation
2
Soundness
2
Read review
Paper summary

This paper presents GeoCFR, a coarse-to-fine framework for multimodal remote sensing image registration. It first estimates a rough alignment using structural features and then refines the global affine transformation by combining information from several possible local matches. The authors also introduce WH-M3R, a dataset covering diverse remote sensing data types and spatial resolutions to support multimodal registration research. Experiments show improved registration accuracy over the evaluated baselines and promising generalization to external datasets.

Critical strengths
  1. GeoCFR outperforms the evaluated baselines across different transformation settings and shows promising cross-dataset generalization.
  2. The dataset is another contribution. WH-M3R covers diverse remote sensing data types and spatial resolutions, providing a useful resource for multimodal registration research.
Critical weaknesses
  1. The authors mentioned CRFT, MINIMA and RoMa in the paper, but these cross-modal image matching methods were not included in Table 4 and 5. Why?
  2. The training benefits of WH-M3R dataset are not clearly presented. Under the same training conditions, does the same network achieve higher registration accuracy and better generalization when trained on WH-M3R than on existing datasets?
Questions for authors

The paper states, “All learning-based methods are retrained on WH-M3R-Match-30K with 512 × 512 inputs.” Are the learning-based baselines trained from scratch or fine-tuned from their respective publicly available pretrained weights? Do all baselines follow the training strategy described in Section 5.1, or does each method retain its original training strategy?

4 comments
AuthorsOpenReview ↗

Overall response

Thank you for raising questions about baseline coverage, dataset training benefits, and initialization. We added results for RoMa trained on WH-M3R and for MINIMA-RoMa using its official weights directly. GeoCFR achieves higher CMR than all four added outputs under the three original protocols and on the three external benchmarks. The comparison using the same WAFT architecture also shows gains on WH-M3R matching and optical–infrared transfer. We explain CRFT's high-resolution adaptation and distinguish component pretraining from task training under the common curriculum below.

AuthorsOpenReview ↗

W1: Additional comparisons with CRFT, MINIMA, and RoMa

We added RoMa and MINIMA-RoMa under the same evaluation protocol, and the comparisons support GeoCFR's original performance conclusions. RoMa was trained on WH-M3R under the common curriculum. MINIMA-RoMa uses its official multimodal pretrained weights directly. We report raw and affine-fitted outputs for both methods. Predictions are mapped back to 512×512 image resolution for scoring.

Table S1: Additional results under the original Low/Medium/High protocols (CMR, %).

LowMediumHighMethod1p˜x2p˜x3p˜x1p˜x2p˜x3p˜x1p˜x2p˜x3p˜xRoMa11.2/29.0‾47.7/55.6‾64.8/68.6‾10.1/24.9‾42.2/50.4‾57.5/62.8‾8.7/20.7‾37.1/46.3‾53.3/58.6‾MINIMA-RoMa2.3/17.810.7/33.817.7/41.00.9/5.73.3/14.16.2/17.70.3/1.81.3/4.32.1/6.5GeoCFR\mathbfsf48.5\mathbfsf71.0\mathbfsf80.7\mathbfsf40.2\mathbfsf63.5\mathbfsf74.1\mathbfsf34.3\mathbfsf57.0\mathbfsf68.7{\small\def\R#1{\mathsf{#1}}\def\B#1{\mathbfsf{#1}}\begin{array}{lccccccccc}\hline&&\mathclap{\begin{array}{c}\R{Low}\\\\[-.65em]\rule{14.6em}{.4pt}\end{array}}&&&\mathclap{\begin{array}{c}\R{Medium}\\\\[-.65em]\rule{14.6em}{.4pt}\end{array}}&&&\mathclap{\begin{array}{c}\R{High}\\\\[-.65em]\rule{14.0em}{.4pt}\end{array}}&\\\\\R{Method}&\R{1\~px}&\R{2\~px}&\R{3\~px}&\R{1\~px}&\R{2\~px}&\R{3\~px}&\R{1\~px}&\R{2\~px}&\R{3\~px}\\\\\hline\R{RoMa}&\sf11.2/\underline{29.0}&\sf47.7/\underline{55.6}&\sf64.8/\underline{68.6}&\sf10.1/\underline{24.9}&\sf42.2/\underline{50.4}&\sf57.5/\underline{62.8}&\sf8.7/\underline{20.7}&\sf37.1/\underline{46.3}&\sf53.3/\underline{58.6}\\\\\R{MINIMA\text{-}RoMa}&\sf2.3/17.8&\sf10.7/33.8&\sf17.7/41.0&\sf0.9/5.7&\sf3.3/14.1&\sf6.2/17.7&\sf0.3/1.8&\sf1.3/4.3&\sf2.1/6.5\\\\\R{GeoCFR}&\B{48.5}&\B{71.0}&\B{80.7}&\B{40.2}&\B{63.5}&\B{74.1}&\B{34.3}&\B{57.0}&\B{68.7}\\\\\hline\end{array}}

\stylefont−size:12pxNote: Original and affine-fitted outputs appear before and after the slash. Rankings use each method’s better output. Bold and\stylefont−size:12pxunderlining mark the best and second-best methods, respectively.{\def\N#1{\style{font-size:12px}{\textsf{#1}}}\begin{array}{l}\N{Note: Original and affine-fitted outputs appear before and after the slash. Rankings use each method's better output. Bold and}\\\\\N{underlining mark the best and second-best methods, respectively.}\end{array}}

Table S2: High-protocol results on the three external benchmarks (CMR, %).

SOMA-1M-TestRoadSceneOSdatasetMethod1p˜x2p˜x3p˜x1p˜x2p˜x3p˜x1p˜x2p˜x3p˜xRoMa0.60/4.30‾11.00/19.30‾25.60/31.40‾0.00‾/0.00‾1.36/3.175.88/13.12‾0.00/0.24‾1.89/7.08‾14.15/21.93‾MINIMA-RoMa0.10/0.400.40/1.100.70/2.000.00‾/0.00‾0.45/3.62‾1.81/7.240.00/0.000.00/0.000.00/0.47GeoCFR\mathbfsf13.90\mathbfsf31.30\mathbfsf41.00\mathbfsf0.90\mathbfsf10.41\mathbfsf24.89\mathbfsf1.89\mathbfsf12.97\mathbfsf30.19{\small\def\R#1{\mathsf{#1}}\def\B#1{\mathbfsf{#1}}\begin{array}{lccccccccc}\hline&&\mathclap{\begin{array}{c}\R{SOMA\text{-}1M\text{-}Test}\\\\[-.65em]\rule{16.6em}{.4pt}\end{array}}&&&\mathclap{\begin{array}{c}\R{RoadScene}\\\\[-.65em]\rule{15.1em}{.4pt}\end{array}}&&&\mathclap{\begin{array}{c}\R{OSdataset}\\\\[-.65em]\rule{15.5em}{.4pt}\end{array}}&\\\\\R{Method}&\R{1\~px}&\R{2\~px}&\R{3\~px}&\R{1\~px}&\R{2\~px}&\R{3\~px}&\R{1\~px}&\R{2\~px}&\R{3\~px}\\\\\hline\R{RoMa}&\sf0.60/\underline{4.30}&\sf11.00/\underline{19.30}&\sf25.60/\underline{31.40}&\sf\underline{0.00}/\underline{0.00}&\sf1.36/3.17&\sf5.88/\underline{13.12}&\sf0.00/\underline{0.24}&\sf1.89/\underline{7.08}&\sf14.15/\underline{21.93}\\\\\R{MINIMA\text{-}RoMa}&\sf0.10/0.40&\sf0.40/1.10&\sf0.70/2.00&\sf\underline{0.00}/\underline{0.00}&\sf0.45/\underline{3.62}&\sf1.81/7.24&\sf0.00/0.00&\sf0.00/0.00&\sf0.00/0.47\\\\\R{GeoCFR}&\B{13.90}&\B{31.30}&\B{41.00}&\B{0.90}&\B{10.41}&\B{24.89}&\B{1.89}&\B{12.97}&\B{30.19}\\\\\hline\end{array}}

\stylefont−size:12pxNote: Original and affine-fitted outputs appear before and after the slash. Rankings use each method’s better output. Bold and\stylefont−size:12pxunderlining mark the best and second-best methods, respectively.{\def\N#1{\style{font-size:12px}{\textsf{#1}}}\begin{array}{l}\N{Note: Original and affine-fitted outputs appear before and after the slash. Rankings use each method's better output. Bold and}\\\\\N{underlining mark the best and second-best methods, respectively.}\end{array}}

GeoCFR exceeds all four added outputs under every protocol and at every threshold reported above. Under High, its CMR@1/2/3 exceeds RoMa+Affine by 13.6/10.7/10.1 percentage points. Its CMR@3 margins over RoMa+Affine on the three external benchmarks are 9.60/11.77/8.26 percentage points, respectively. These comparisons cover both task training on WH-M3R and direct transfer using publicly released multimodal weights.

We trained CRFT with the original paper's 64×64 patch setting during preparation of the original submission. We also tested mapping its predictions back to 512×512. Directly expanding the current implementation to 512×512 required excessive memory. The mapped low-resolution predictions did not meet our strict pixel-accuracy thresholds, so this adaptation was omitted from the original Tables 4 and 5. We further examined the memory requirements at 512×512 with FP32 and batch size 1. A single attention-score tensor requires 87.50 GiB, exceeding the capacity of one A100 80GB. We report the completed low-resolution adaptation below.

Table S4: CRFT low-resolution adaptation results (64×64 training, with predictions mapped to 512×512 for evaluation).

{\small\def\R#1{\mathsf{#1}}\begin{array}{lccccc}\hline\R{Protocol}&\R{CMR@1\~(\\%)}&\R{CMR@2\~(\\%)}&\R{CMR@3\~(\\%)}&\R{CMR@5\~(\\%)}&\R{EPE\~(px)}\\\\\hline\R{Low}&\sf0.0&\sf0.0&\sf0.1&\sf1.8&\sf22.450\\\\\R{Medium}&\sf0.0&\sf0.0&\sf0.0&\sf0.2&\sf51.228\\\\\R{High}&\sf0.0&\sf0.0&\sf0.0&\sf0.0&\sf122.704\\\\\hline\end{array}}

\stylefont−size:12pxNote: Each protocol contains 1,000 image pairs and uses the model trained for that protocol. CMR is computed per image pair; EPE is\stylefont−size:12pxweighted by the number of valid pixels.{\def\N#1{\style{font-size:12px}{\textsf{#1}}}\begin{array}{l}\N{Note: Each protocol contains 1,000 image pairs and uses the model trained for that protocol. CMR is computed per image pair; EPE is}\\\\\N{weighted by the number of valid pixels.}\end{array}}

Revised sections: pp. 7–9, Sections 5.1–5.2 (Experimental Setup; Results), Tables 4–5 and Figure 3.

AuthorsOpenReview ↗

W2: Training and cross-modal transfer benefits of WH-M3R

WH-M3R training benefits both registration across data layers and transfer to other multimodal scenes. We ran training experiments using three existing optical–SAR datasets: SOMA, 3MOS, and GUSO. We fixed WAFT's architecture, component initialization rules, optimization settings, and Low→Medium→High curriculum. 3MOS and GUSO did not converge in these training runs. The table compares the SOMA model trained through Low, Medium, and High with the WH-M3R-trained model on four fixed High-protocol test sets.

Table S3: Registration and cross-modal transfer of WAFT+Affine trained on different datasets (High, CMR@1/2/3, %). Δ is WH-M3R training minus SOMA training in CMR@3.

Tests˜etWH-M3Rt˜rainingSOMAt˜rainingΔCMR@3(˜pp)WH-M3R\mathbfsf26.60/\mathbfsf50.10/\mathbfsf59.8013.70/40.40/50.00+9.80RoadScene0.00/\mathbfsf8.60/\mathbfsf24.890.00/4.98/17.65+7.24SOMA-1M-Test\mathbfsf10.60/28.20/37.2010.20/\mathbfsf35.40/\mathbfsf49.00−11.80OSdataset\mathbfsf3.07/11.08/27.362.59/\mathbfsf12.03/\mathbfsf31.60−4.24{\small\def\R#1{\mathsf{#1}}\def\B#1{\mathbfsf{#1}}\begin{array}{lccc}\hline\R{Test\~set}&\R{WH\text{-}M3R\~training}&\R{SOMA\~training}&\R{ΔCMR@3\~(pp)}\\\\\hline\R{WH\text{-}M3R}&\sf\B{26.60}/\B{50.10}/\B{59.80}&\sf13.70/40.40/50.00&\sf+9.80\\\\\R{RoadScene}&\sf0.00/\B{8.60}/\B{24.89}&\sf0.00/4.98/17.65&\sf+7.24\\\\\R{SOMA\text{-}1M\text{-}Test}&\sf\B{10.60}/28.20/37.20&\sf10.20/\B{35.40}/\B{49.00}&\sf-11.80\\\\\R{OSdataset}&\sf\B{3.07}/11.08/27.36&\sf2.59/\B{12.03}/\B{31.60}&\sf-4.24\\\\\hline\end{array}}

Multimodal accuracy and external transfer. On WH-M3R-Match-1K, which includes 10 target data layers, WH-M3R training raises CMR@3 from 50.00% to 59.80% (+9.80 percentage points). On the external optical–infrared benchmark RoadScene, it raises CMR@3 from 17.65% to 24.89% (+7.24 percentage points). Raw flow shows the same benefit. In these two tests, CMR@3 increases from 48.80%/14.03% to 58.30%/20.81%, respectively.

Training coverage and generalization. SOMA training yields higher CMR@2/3 on SOMA-1M-Test and OSdataset, consistent with specialization for optical–SAR pairs. Even on these two optical–SAR benchmarks, WH-M3R training gives slightly higher CMR@1 (10.60% vs. 10.20%; 3.07% vs. 2.59%).

Additional response material: Table S3 (WH-M3R versus SOMA training-source comparison, provided in this response).

AuthorsOpenReview ↗

Q1: Component initialization and the common task-training strategy

The learning-based methods in the original main comparison follow our common task-training strategy while retaining their required component initialization. Retraining on WH-M3R therefore does not mean randomly initializing every parameter.

For methods with pretrained components, we retain their original initialization rules before task training on WH-M3R. For example, WAFT uses a publicly pretrained, frozen Depth Anything V2 encoder. Its refinement and related feature layers follow the original method's initialization and continue training. These methods follow the common Low→Medium→High curriculum and are evaluated using the model for each corresponding protocol.

RoMa loads pretrained DINOv2 weights and freezes this backbone during WH-M3R task training, following its official configuration. Its remaining trainable components undergo task training on WH-M3R. MINIMA-RoMa instead uses its official weights directly for testing. The two weight sources are stated separately in the tables. Both methods run internally at 560×560 and map their outputs back to 512×512 for scoring.

Revised sections: pp. 7–8, Section 5.1 (Experimental Setup), “Training strategy” and “Baselines” paragraphs.

Comments on the paper

2 comments
AuthorsOpenReview ↗

General Response and Summary of Revisions

We thank the reviewers and Area Chair for their careful reading and constructive comments. We have added the following experiments and clarifications.

  1. Baseline comparisons. We added raw and affine-fitted results for RoMa and MINIMA-RoMa. GeoCFR still leads all nine CMR settings under the original protocols (Table 4). The reason for excluding CRFT from the main tables is explained in our response to Reviewer Sja7 W1 (Table S4).
  2. Training benefits of WH-M3R. Using the same WAFT architecture, we compared WH-M3R and SOMA training. For WAFT+Affine, WH-M3R training improves CMR@3 by 9.80 and 7.24 percentage points on WH-M3R-Match-1K and RoadScene, respectively. Results appear in our reply to Reviewer Sja7.
  3. General affine evaluation. We added shear and anisotropic scaling tests without further training. GeoCFR remains competitive overall under these general affine conditions (Appendix H, Tables 18–19).
  4. Evaluation reliability. We added data-layer exclusion, Map/DEM comparisons, and label-translation sensitivity tests. Under High, GeoCFR retains its advantage after excluding Depth, Map, and DEM. Its CMR@1 lead also persists under the tested label shifts up to 1 pixel (Appendices A.2.1 and E).
  5. Posterior information versus capacity. With the PATAR architecture and parameter count fixed, the full posterior outperforms statistics-only encoding in all nine settings (Table 6). Changes in the revised manuscript are marked in blue. Detailed results and responses are provided separately for each reviewer.
Reviewer_mTQsOpenReview ↗

Thank you for the author's reply. I will maintain my original opinion.

Thank you for the author's reply. I will maintain my original opinion.

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.