acceptodds
Under review as a conference paper at ICLR 2027

AeroSCR: Recovering Stable Scene-Coordinate Decodability Under Scale and Rotation Shifts

Abstract

Recent scene coordinate regression (SCR) methods have narrowed the robustness gap to structure-based localization through stronger local representations and improved generalization. Yet SCR remains vulnerable when query observations differ substantially from the mapping conditions. We investigate this gap in near-nadir aerial relocalization, where altitude and heading changes produce large shifts in image scale and in-plane orientation. Controlled experiments show that descriptor variation can preserve relative correspondence discrimination while destabilizing scene-coordinate predictions, revealing different robustness requirements for correspondence retrieval and coordinate decoding. Motivated by this distinction, AeroSCR expands decoder training coverage across scale changes, canonicalizes large heading changes in descriptor space, and learns tolerance to residual yaw errors. Direct metric supervision and retrieval-gated GeoCells further adapt SCR to weak aerial geometry and large-area maps. Across six scenes from AnyVisLoc, AeroSCR reaches 90.8% success at 5 m/5, essentially matching HLoc at 90.2%, with 4.0 less scene-specific storage and 8.2 lower mean online latency. On the roughly 10 km Oslo-Large scene, it achieves 86.2% success with 6.0 less storage and 3.1 lower latency than HLoc.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
scene coordinate regression, visual localization, representation learning, rotation equivariance, scale robustness, aerial relocalization

Reviews

Among:
Shape:
1234this paper, Top 49%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.33 over 3 reviews, above 51% of the others.
Reviewer_4QNaOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper investigates the sensitivity of scene coordinate regression (SCR) to scale and rotation changes in aerial visual localization. The authors observe that local descriptors may retain strong correspondence discrimination while failing to support stable 3D scene-coordinate decoding. To address this issue, AeroSCR combines scale augmentation, descriptor-space rotation canonicalization, residual yaw training, direct coordinate supervision, and GeoCell-based routing. Experiments show that AeroSCR achieves localization success rates comparable to HLoc at the 5 m/5° threshold, while reducing scene-specific storage and inference latency.

Critical strengths
  1. Through controlled experiments, the paper reveals a distinction between local descriptor correspondence discrimination and stable scene-coordinate decoding, providing a valuable perspective on understanding SCR robustness under scale and rotation changes.
Critical weaknesses
  1. Lack of qualitative evaluation. The paper does not present any qualitative localization results on actual image pairs. Real-image visualizations are limited to dataset examples and method illustrations. Visualizations of predicted scene coordinates, localization results, success and failure cases, and qualitative comparisons with baseline methods on the same images are entirely absent.

  2. Insufficient comparisons with recent methods. The experiments mainly compare against methods from 2025 or earlier, lacking comparisons with representative methods from 2026. The limited baseline coverage makes it difficult to establish the method's competitiveness against recent state-of-the-art approaches, leaving the experimental evaluation insufficient.

  3. Limited methodological novelty. The paper lacks clear independent algorithmic novelty and mainly combines existing techniques for aerial visual localization. Scale augmentation relies on conventional cropping and resizing, rotation canonicalization uses pretrained and frozen SO(2) Steerers, and both direct coordinate supervision and GeoCell routing build upon existing research.

Questions for authors
  1. Performance gap under stricter localization thresholds.

AeroSCR achieves performance comparable to HLoc on AnyVisLoc at the 5 m/5° threshold (90.8% vs. 90.2%), but falls substantially behind at 1 m/1° (60.4% vs. 80.3%). A similarly large gap appears on Oslo-Large at 2 m/2° (43.2% vs. 76.3%).

What are the main reasons for these performance gaps? Do these results indicate that AeroSCR primarily improves localization robustness under relatively loose thresholds, while its high-precision localization capability remains limited?

4 comments

Thank you for your constructive feedback and for recognizing the value of our controlled analysis distinguishing correspondence discrimination from coordinate decoding. We address each of your points below.

To CW1 (Qualitative evaluation):

We have added qualitative visualizations across real-world (AnyVisLoc, AGL/AMSL) and controlled (Oslo-Center, Oslo-Large) scenes, covering baseline comparisons on identical query views, single-component ablations, and representative failure cases (see Sec. 5.7, Figs. 4–5 in the revised main text and Appendix G.2, Fig. 9). By projecting the full sparse 3D point clouds predicted by SCR methods into the ground-truth query view and coloring points by scene elevation (ZZ), these visualizations show that AeroSCR preserves coherent building and terrain structure under large scale and heading shifts where native R-SCoRe scatters, while illustrating the complementary roles of scale augmentation and residual-yaw training.

To CW2 (Comparisons with recent 2026 methods):

We added experimental comparisons with Robust SCR (Nguyen et al., WACV 2026), a recent R-SCoRe extension that learns geometrically consistent global descriptors, evaluated across AGL/AMSL, all six AnyVisLoc scenes, and Oslo-Center using the same mapping-only mesh targets (updated Tables 2–4 and Appendix D.5, Table 11). While Robust SCR performs well under fixed heading on Oslo-Center (91.1% at @5m/5° and 66.2% at @2m/2°), its upright representation remains vulnerable to large aerial yaw shifts, dropping to 13.6% on AGL/AMSL (vs. AeroSCR's 100.0%). Across the six AnyVisLoc scenes, AeroSCR outperforms Robust SCR at all thresholds, reaching 90.8% vs. 56.3% at @5m/5° and 60.4% vs. 29.4% at @1m/1°, while requiring 4.1× less storage (10.6 MB vs. 43.4 MB) and 5.2× lower latency (111 ms vs. 576 ms).

Regarding other 2026 works cited in Sec. 2, complete official code and pretrained retrieval-descriptor weights for CoLoR (CVPR 2026) are not yet publicly available, while Triplet-based map compression (CVPR 2026) and AsymLoc (CVPR 2026) focus on explicit SfM map compression or detector-descriptor distillation rather than implicit scene-coordinate decoding.

To CW3 (Clarifying contributions beyond combining existing techniques):

Rather than an ad-hoc combination of existing primitives, our contribution lies in the empirical insight, the resulting decoder-side supervision paradigm, and the system design that makes compact SCR viable under large aerial observation shifts:

  1. Decoder-side cross-view consistency. Our controlled probes show that scene-coordinate decoding deteriorates sharply under scale and rotation shifts even when local descriptors retain strong correspondence discrimination (Fig. 1, Sec. 3, and Appendix C). Unlike methods promoting cross-view consistency via feature-level invariance (which risks discarding orientation-sensitive cues and reducing fine discriminability), we preserve informative, transformation-sensitive descriptors and establish coordinate consistency through decoder supervision (Sec. 1, Sec. 2.2).
  2. Asymmetric bounded-residual supervision. Requiring a compact decoder to absorb full-360° rotations from limited mapping views strains model capacity (Fig. 1b; Appendix C), whereas unaligned mapping headings weaken multi-view consistency. We align observations to a canonical heading via steerable descriptors and restrict coordinate supervision strictly to a narrow residual neighborhood [−ϵmax⁡,+ϵmax⁡][-\epsilon_{\max}, +\epsilon_{\max}], deliberately leaving larger rotations unsupervised (Sec. 4.3). While discrete steering or residual training alone is insufficient due to quantization gaps or limited rotation coverage (Table 5b), our asymmetric supervision provides local rotation tolerance while giving misaligned yaw hypotheses weaker geometric support during RANSAC verification. This enables 98.1% prior-free full-circle localization using standard P3P without any yaw or gravity prior (Appendix D.1, Table 7a).
  3. Making near-nadir aerial SCR viable at scale and benchmark contribution. Existing SCR methods designed for upright ground-level scenes struggle under aerial viewpoint shifts and weak projective depth geometry. We address these domain bottlenecks through coverage-balanced scale sampling across altitude extremes (Fig. 3), direct metric supervision for weak near-nadir depth conditioning (Table 5a), and GeoCell routing to avoid the convergence failure of shared global-conditioned decoders on kilometer-scale maps (Appendix D.4). Together with our tailored evaluation protocols on public UAV datasets and the controlled Oslo-Center and Oslo-Large (~10 km²) benchmarks for public release, these designs establish a practical framework for aerial SCR.

To Q1 (Strict-threshold precision and high-precision capability):

1. Reasons for the gap to HLoc at the strictest thresholds:

  • Implicit neural compression vs. explicit triangulated point clouds: HLoc stores explicit 3D SfM points and descriptors, estimating poses from exact triangulated feature tracks. In contrast, AeroSCR compresses scene geometry into compact MLP heads that regress continuous 3D coordinates. At high flight altitudes (300–1000 m on Oslo-Center and Oslo-Large), continuous neural regression introduces ~1–2 m smoothing errors that have minor impact at @5m/5° (86.2% vs. HLoc's 90.5% on Oslo-Large) but cross the strict @1m or @2m cutoffs, reflecting the trade-off for 4.0–6.0× smaller storage and 3.1–8.2× lower latency.
  • Approximate view-shift modeling and local invariance trade-off: To enable fast single-pass feature extraction at inference, AeroSCR approximates 3D altitude changes via 2D crop-resize augmentation and handles large heading shifts via approximate SO(2) descriptor steering with ±15∘\pm 15^\circ residual tolerance. As shown in our residual-radius sensitivity analysis (Appendix D.3, Table 9), broader local rotational tolerance slightly softens peak coordinate precision at @1m/1°.

2. High-precision capability of AeroSCR: These trade-offs do not restrict AeroSCR to coarse localization. At lower flight altitudes (AGL/AMSL, Table 2), AeroSCR achieves 99.5% at @1m/1° (0.15 m / 0.15° median error), nearly matching HLoc (100.0%). Across the six AnyVisLoc scenes at @1m/1° (Table 3), AeroSCR (60.4%) outperforms R-SCoRe (36.3%) and Robust SCR (29.4%) by 24.1 and 31.0 percentage points, substantially advancing fine-grained SCR accuracy while retaining compact storage and fast online inference.

Area_Chair_wdoWOpenReview ↗

Invitation to Review and Discuss the Rebuttal

Dear Reviewer,

Thank you for your valuable feedback on this submission. Since we are now in the discussion phase, I would be grateful if you could engage with the authors’ responses and continue the conversation if there are outstanding issues that would benefit from clarification. Your participation will help ensure a constructive and thorough evaluation.

Many thanks again for your time and consideration.

Best regards,

Area Chair

Thank you for the rebuttal and clarifications. I have reviewed the authors' responses and will maintain my original rating.

Reviewer_Sd5dOpenReview ↗
Rating
2
Confidence
2
Informativeness
1
Presentation
2
Soundness
2
Read review
Paper summary

This paper addresses UAV image relocalization using scene coordinate regression. It shows that descriptors can remain reliable for matching while producing unstable coordinate predictions. It proposes AeroSCR, a training framework that combines scale augmentation, descriptor-space rotation canonicalization, and residual rotation training. Experiments show better robustness than SCR baselines and competitive localization performance compared with HLoc, with lower scene-specific storage requirements and faster inference. Still, HLoc remains more accurate at stricter localization thresholds.

Critical strengths
  • The analysis of the different descriptor requirements for feature matching and scene coordinate regression is interesting.

  • AeroSCR achieves comparable success to HLoc at 5 m/5° while using approximately 4× less scene-specific storage.

  • Ablation study shows clear benefit of the introduced augmentation, metric supervision, yaw canonicalization and residual training.

Critical weaknesses
  • In general, I am not convinced by the use of SCR for UAV relocalization. Since near-nadir drone imagery can often be approximated by weak-perspective projection, predicting full 3D coordinates appears weakly constrained, as variations in scene height (Z) have little effect on the projected image coordinates.

  • Although the analysis of the different descriptor requirements for feature matching and scene coordinate regression is convincing, I find the methodological contribution relatively incremental. The proposed solution mainly combines scale augmentation, pretrained descriptor steering, and residual rotation augmentation, and its novelty beyond adapting and combining these existing techniques is unclear.

  • Seems there is no study on how optimizing for matching impacts SCR, or vice versa.

Questions for authors
  • Given the weak depth constraints in near-nadir imagery, how accurately does AeroSCR predict scene height (Z), and how sensitive is localization accuracy to errors in this coordinate?
6 comments

Thank you for your thoughtful review and for finding our analysis of the distinct requirements for feature matching and scene-coordinate decoding convincing. We address your concerns below.

To W1 and Q1 (Weak depth constraints in near-nadir imagery, accuracy of predicted ZZ, and localization sensitivity):

We thank the reviewer for this insightful geometric observation. Under near-nadir projection, 2D reprojection error alone indeed provides weak constraints on scene height (ZZ). For this reason, AeroSCR incorporates direct metric coordinate supervision from the mapping mesh (Sec. 4.4, Eq. 10) rather than relying solely on reprojection. On Oslo-Center, removing this metric loss reduces @2 m/2∘2\text{ m}/2^\circ localization success from 78.1% to 69.5% (Table 5a), confirming its role under weak projective depth conditioning. To directly answer Q1, we evaluated the frozen Oslo-Center model across all 370370 fixed-heading queries:

1. Accuracy of Predicted Scene Height (ZZ).

  • Quantitative accuracy: Evaluated against the mapping mesh on Oslo-Center, refined predictions achieve a median absolute ZZ error of 1.86 m1.86\text{ m} among pose-supported inliers (65.6% of points) and 2.87 m2.87\text{ m} across all valid predictions, representing less than 0.7% of the 300–1000 m300\text{--}1000\text{ m} flight altitude. On lower-altitude real UAV datasets, the inlier median ZZ error is 0.11 m0.11\text{ m} on AGL/AMSL and 0.67 m0.67\text{ m} across the six AnyVisLoc scenes (Appendix G.3, Table 13).
  • Qualitative evidence: As shown in our added elevation-colored 3D point cloud visualizations (Sec. 5.7, Figs. 4–5 in the revised main text), the predicted ZZ coordinates capture 3D building roofs and terrain elevation changes in close agreement with the mesh reference.

2. Sensitivity of Localization to ZZ Errors. Holding predicted (X^,Y^)(\hat{X}, \hat{Y}) and 2D keypoint locations fixed across all 5,0005,000 correspondences per query on Oslo-Center, we perturbed only the predicted Z^\hat{Z} passed to the pose solver with independent Gaussian noise N(0,σZ2)\mathcal{N}(0, \sigma_Z^2) (Appendix G.3, Table 14):

ZZ-Only Perturbation (Oslo-Center)Original+N(0,0.52)+\mathcal{N}(0, 0.5^2)+N(0,12)+\mathcal{N}(0, 1^2)+N(0,22)+\mathcal{N}(0, 2^2)+N(0,52)+\mathcal{N}(0, 5^2)+N(0,102)+\mathcal{N}(0, 10^2)
Success @5 m/5∘5\text{ m}/5^\circ (%)99.799.799.799.796.970.4
Success @2 m/2∘2\text{ m}/2^\circ (%)78.176.976.173.854.822.9
  • AeroSCR remains robust to moderate vertical noise: for perturbations up to σZ=2 m\sigma_Z = 2\text{ m} (above its 1.86 m1.86\text{ m} median inlier error), multi-point geometric consensus preserves 99.7% success at @5 m/5∘5\text{ m}/5^\circ and 73.8% at @2 m/2∘2\text{ m}/2^\circ, with localization accuracy degrading substantially only under severe noise (σZ=5–10 m\sigma_Z = 5\text{--}10\text{ m}).

To W2 (Clarifying contributions beyond adapting and combining existing techniques):

We appreciate the opportunity to clarify the value of our work. Rather than proposing isolated low-level operators, our contribution lies in the empirical insight, the resulting decoder-side supervision paradigm, and the domain-specific system design that makes compact SCR viable for cross-altitude, cross-heading aerial relocalization:

  1. From empirical finding to asymmetric bounded-residual supervision. Building directly on the distinction between relative correspondence discrimination and absolute coordinate decoding (Fig. 1, Sec. 3, Appendix C), we preserve informative, transformation-sensitive descriptors rather than enforcing feature-level invariance, and establish cross-view consistency through decoder supervision (Sec. 1). Because forcing a compact decoder to absorb full-360∘360^\circ rotations from limited mapping views strains model capacity (Fig. 1b), we introduce asymmetric bounded-residual supervision (Sec. 4.3): we align observations to a canonical heading via steerable descriptors and restrict coordinate supervision strictly to a narrow residual neighborhood [−ϵmax⁡,+ϵmax⁡][-\epsilon_{\max}, +\epsilon_{\max}], leaving larger rotations unsupervised. While discrete steering or residual training alone is insufficient due to quantization gaps or limited rotation coverage (Table 5b), our asymmetric supervision provides local rotation tolerance while implicitly suppressing misaligned yaw hypotheses during RANSAC verification, enabling 98.1% prior-free full-circle localization with standard P3P (Appendix D.1, Table 7a).
  2. Making near-nadir aerial SCR viable where upright ground assumptions fail. Existing SCR methods designed for upright ground-level scenes struggle under large aerial observation shifts. We identify and address three essential prerequisites overlooked in ground-level settings: (i) direct metric coordinate supervision to overcome weak near-nadir projective depth conditioning (Table 5a); (ii) coverage-balanced scale sampling to prevent degradation at extreme flight altitudes (Fig. 3); and (iii) retrieval-gated GeoCells to avoid the optimization collapse of shared global-conditioned decoders on kilometer-scale maps (Appendix D.4). Together, these designs turn aerial SCR into a practical alternative that matches HLoc at @5 m/5∘5\text{ m}/5^\circ with 4.0–6.0×4.0\text{--}6.0\times less storage and 3.1–8.2×3.1\text{--}8.2\times lower latency.
  3. New aerial relocalization task setting and public benchmarks. We formulate near-nadir 6-DoF aerial relocalization under substantial query scale and yaw shifts from constrained mapping views, establishing tailored evaluation protocols on public UAV datasets and constructing the controlled Oslo-Center and kilometer-scale Oslo-Large (∼10 km2\sim 10\text{ km}^2) benchmarks for public release.

To W3 (How optimizing for matching impacts SCR, and vice versa):

We address both directions of this question below:

  1. How optimizing for feature matching impacts SCR. Descriptors trained for feature matching optimize relative local distinctiveness rather than strict transformation invariance (Sec. 3.1). Our controlled probes (Fig. 1, Appendix C) show that matching-optimized descriptors (DeDoDe-B) remain highly discriminative under moderate scale and yaw shifts, whereas a fixed SCR decoder collapses once descriptor variations move queries outside its training support (Sec. 3.2). Since enforcing broad invariance in descriptor space discards orientation-sensitive cues and harms fine discriminability (Sec. 1), preserving transformation-sensitive matching descriptors is beneficial for SCR once paired with canonical alignment. Conversely, forcing the SCR decoder to absorb broad rotational invariance dilutes its capacity and degrades precision:
    • In our augmented yaw probe (Fig. 1b; Appendix C), direct RGB rotation augmentation with prolonged optimization still leaves a wide gap to explicit NN matching (29.3% vs. 62.1% mean success at @5 m/5∘5\text{ m}/5^\circ).
    • In Appendix D.3 (Table 9), widening the residual-supervision radius ϵmax⁡\epsilon_{\max} beyond moderate values consistently reduces strict @1 m/1∘1\text{ m}/1^\circ accuracy across evaluation protocols, confirming the trade-off between broad rotational tolerance and fine coordinate precision.
  2. How optimizing for SCR impacts feature matching. AeroSCR keeps the pretrained feature extractor frozen and estimates poses via direct coordinate regression without online descriptor matching. Evaluating the reverse impact of scene-coordinate supervision on explicit feature matching would require fine-tuning the descriptor backbone, which is outside the scope of this work but remains an interesting direction for future study.
Area_Chair_wdoWOpenReview ↗

Invitation to Review and Discuss the Rebuttal

Dear Reviewers,

Thank you again for your thoughtful reviews and the time you have devoted to this submission. As we are now in the discussion phase, I would like to kindly encourage you to engage with the authors’ responses and continue the discussion where appropriate by sharing any remaining questions, clarifications, or concerns.

Your participation will be greatly appreciated and will help ensure a constructive and well-informed final evaluation.

Thank you very much for your continued effort and consideration.

Best regards,

Area Chair

Reviewer_FN2sOpenReview ↗
Rating
3(+1)
Confidence
4
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper aims to improve the robustness of Scene Coordinate Regression (SCR) Localization on aerial images with significant scale and rotation variations. It demonstrate the generalization issue of SCR to unseen scale and rotation and propose to mitigate this problem with dedicated augmentation and canonicalization for scale and rotation. Its evaluation shows comparable accuracy to HLoc with better runtime and map size efficiency.

Critical strengths
  1. Practical and well motivated. Aerial images usually have larger scale and rotation variations than ground images. Better augmentation and canonicalization strategy is a practical way to solve this issue.
  2. Honest evaluation. Strong baselines including HLoc and R-SCoRe is included, and added a R-SCoRe+ baseline with comparable training setting.
Critical weaknesses
  1. Limited novelty. Data augmentation for scale and rotation is already widely used in previous SCR works. While strategies like coverage-balanced sampling and canonicalization with descriptor steering dedicated to aerial images with large scale and rotation changes are practical and well motivated improvements, it still seems incremental. Also for the analysis in Figure 1, the SCR head is trained without scale/rotation augmentation, which is not representative for the latest best practice for SCR training.
  2. Limited comparison for baselines with orientation prior. The main results for AeroSCR still require orientation prior, while most baselines like ACE-G and GLACE are still evaluated without it. Also not enough evaluation for AeroSCR without orientation prior to see how well it can really handle large rotation variation itself without additional cues.
Questions for authors
  1. For Figure 1, what if we apply sufficient data augmentation to the SCR head training?
  2. For the SCR baselines, what if we apply the orientation prior to canonicalize the image?
  3. Are there any comparisons about the pose solver?
4 comments

Thank you for your detailed review and constructive suggestions. We have conducted additional experiments and clarified several aspects of our work to address your concerns.

To W1 and Q1:

1. Scale and Rotation Augmentation in Figure 1. We thank the reviewer for highlighting that the SCR head in the original Figure 1 was trained without scale or rotation augmentation. Figure 1 examines how feature matching and SCR, using the same frozen DeDoDe-B representation, differ as query conditions extend beyond mapping views. We repeated both controlled probes with extensive image-level augmentation (9 augmented exposures per mapping view) and 10× longer training (300k vs. 30k steps; see updated Fig. 1 and Appendix C):

5m/5° (%)SCRSCR+augNN
Scale augN/A9×LogUniform[0.8,1.3]N/A
Scale10.477.3100.0
Rotation augN/A9×Uniform[0°,90°]N/A
Rotation9.329.362.1

Augmentation substantially improves SCR robustness under scale changes, though noticeable degradation persists at the scale boundaries (57.1% at 85% and 120% relative scale). Direct RGB rotation augmentation brings much more limited gains (29.3% mean success), leaving a wide gap to NN matching (62.1%) and motivating AeroSCR's canonical steering with bounded-residual supervision.

2. Novelty and Significance. We highlight three core contributions in the revised manuscript (Sec. 1):

  1. Decoder-side cross-view consistency. We distinguish descriptor-level correspondence discrimination from reliable scene-coordinate decoding (Fig. 1, Sec. 3). Unlike methods promoting cross-view consistency via feature-level invariance, which risks discarding orientation-sensitive cues and reducing fine discriminability, we preserve informative, transformation-sensitive descriptors and establish coordinate consistency via decoder supervision.
  2. Asymmetric bounded-residual supervision. Requiring a compact decoder to absorb full-360° rotations from limited mapping views strains model capacity. We align mapping descriptors to a canonical heading via steerable descriptors and restrict coordinate supervision to a narrow residual yaw range [−ϵmax⁡,+ϵmax⁡][-\epsilon_{\max}, +\epsilon_{\max}], deliberately leaving larger rotations unsupervised (Sec. 4.3). This enables locally rotation-tolerant decoding while giving severely misaligned yaw hypotheses weaker geometric support during RANSAC pose verification (Table 5b), enabling accurate full-circle localization without orientation priors (Appendix D.1, Table 7a).
  3. Cross-altitude and cross-heading aerial benchmarks. We establish a practical near-nadir 6-DoF aerial relocalization setting with limited mapping-view altitude and heading coverage across substantial query-time scale and yaw shifts, introducing tailored evaluation protocols on public UAV datasets and constructing the controlled Oslo-Center and kilometer-scale Oslo-Large (~10 km²) benchmarks for public release.

To W2 and Q2:

1. Applying Orientation Priors to SCR Baselines. We evaluated frozen ACE-G and GLACE models using the same coarse yaw and gravity estimates (±60∘/10∘\pm 60^\circ / 10^\circ) available to AeroSCR (Appendix G.1, Table 12). The yaw estimate guides query-image rotation toward the top-1 retrieved mapping view across 9 candidates searched at 15° intervals, and the gravity estimate is incorporated via UP2P:

5m/5° (%)Native+Yaw+Gravity/UP2P
AGL/AMSL - ACE-G21.843.753.4
AGL/AMSL - GLACE13.196.699.0
AnyVisLoc - ACE-G21.520.420.5
AnyVisLoc - GLACE22.520.923.7

While test-time orientation priors substantially improve both baselines on the single-flight AGL/AMSL dataset (GLACE reaches 99.0% vs. AeroSCR's 100.0%), they bring negligible gains on the more diverse AnyVisLoc benchmark (20.5%–23.7% vs. AeroSCR's 90.8%), confirming that test-time orientation priors alone fail to achieve accuracy comparable to AeroSCR across diverse aerial scenes.

2. AeroSCR without Orientation Priors. We evaluated the frozen AeroSCR model without external yaw or gravity priors on all 370 random-heading queries of Oslo-Center by widening descriptor steering to 24 hypotheses covering the full 360° and replacing UP2P with standard unconstrained P3P (Appendix D.1, Table 7a):

YawGravitySolver#Hyp5m/5° (%)2m/2° (%)
YesYesUP2P997.859.7
NoYesUP2P2498.659.2
NoNoP3P2498.158.9

AeroSCR maintains comparable accuracy without orientation priors, achieving 98.1% at 5m/5° and 58.9% at 2m/2° vs. 97.8% and 59.7% under the default setting, confirming the feasibility of fully prior-free localization.

To Q3: Pose Solver Comparison

In the full-circle Oslo-Center evaluation (Appendix D.1, Table 7a), gravity-constrained UP2P and standard unconstrained P3P (both followed by unconstrained 6-DoF refinement) achieve comparable success rates (98.6% vs. 98.1% at 5m/5°, and 59.2% vs. 58.9% at 2m/2°) on identical correspondences and 24 full-circle hypotheses, confirming that AeroSCR does not rely on minimal-solver gravity constraints for pose accuracy.

Area_Chair_wdoWOpenReview ↗

Invitation to Continue the Discussion

Dear Reviewer,

Thank you again for your thoughtful review and the time you have devoted to this submission. As the discussion period is underway, I would greatly appreciate it if you could consider the authors’ response and engage further in the discussion should you have any remaining questions or concerns. A brief follow-up on whether the response addresses your points would also be very helpful.

Best regards,

Area Chair

Thank authors for their additional experiments and revision. W1 Q1: partially resolved. The comparison to data augmentation better justifies the motivation for canonicalization, and revised core contribution including the new benchmark better clarified the significance. This empirical insight and system design are interesting and useful to certain extent. W2 Q2: mainly resolved. Additional comparisons about effect orientation priors, especially the ones on AnyVisLoc reduce my concerns, and shows the importance of doing canonicalization at training time instead of test-time only. Q3: resolved. Thanks for clarifying this. Overall the critical concerns have been mainly addressed, and the paper provides motivated insights and useful empirical design for certain application though I am still not fully convinced that the contribution is significant enough, so I am raising the score to weak accept.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.