- 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
- The method is conceptually well motivated and addresses key challenges in registration.
- 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
- The current experiments do not sufficiently validate registration under general affine transformations.
- 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
- Can the authors provide experiments with general affine transformations?
- How much do residual alignment errors and differences in data-layer difficulty affect the conclusions?
- How can the gains from PATAR be distinguished from those due to additional model capacity?
4 comments
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, %).
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.
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.
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, %).
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.
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.