acceptodds
Under review as a conference paper at ICLR 2027

Learning Symmetry-Aware RoI Representations for Oriented Object Detection

Abstract

Oriented detection localizes objects with boxes that include orientation, and has found broad applications in embodied intelligence, autonomous driving, and satellite remote sensing. Two-stage oriented detectors typically extract features from rotated proposals, then use these features for classification and box refinement in each proposal's local coordinate frame. However, their detection heads typically do not explicitly account for the symmetry of the proposal's local coordinate frame. Reversing either local axis gives an equivalent description of the same proposal, but affects the prediction targets differently. Class and size remain unchanged, while center offsets and relative angles undergo predictable sign changes. We argue that region features should follow the same transformation laws as the targets they predict. To this end, we propose Symmetry-Aware Region Representations (SRR). We extract features in four equivalent local frames and combine them through the proposed Geometry-Consistent Aggregation (GCA), which weights the frames according to how each target changes sign. We further propose Symmetry-Aware Feature Learning (SFL), which divides features into types by their transformation laws and learns a separate linear map for each type, keeping the head equivariant to local axis reversals. Experiments on DOTA-v1.0, HRSC2016, DIOR-R, RSAR, and FAIR1M show that SRR improves detection performance, with localization gains across multiple detectors and backbones on DOTA-v1.0.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
Oriented object detection, remote sensing images, equivariant representation learning, region representation learning

Reviews

Among:
Shape:
1234this paper, Top 30%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.67 over 3 reviews, above 70% of the others.
Reviewer_EEMTOpenReview ↗
Rating
2
Confidence
4
Informativeness
1
Presentation
3
Soundness
3
Read review
Paper summary

This paper proposes Symmetry-Aware Region Representations (SRR) for two-stage oriented object detection. By modeling local axis reversals as a four-element symmetry group, the proposed GCA module constructs target-specific equivariant features, while SFL learns separate transformations for different feature types. Experiments across five remote sensing benchmarks demonstrate consistent improvements, particularly in localization accuracy.

Critical strengths
  1. The paper introduces a well-motivated symmetry-based formulation for learning region representations in oriented object detection.

  2. The proposed GCA and SFL modules provide a mathematically grounded architecture with explicitly enforced local equivariance.

Critical weaknesses
  1. Although the proposed architecture guarantees equivariance under axis reversals, the paper does not directly establish how this constraint translates into improved robustness or localization. Controlled experiments on equivalent frame representations, proposal perturbations, and different object orientations would better substantiate the central hypothesis.

2.The paper reports improvements in AP75 and average localization accuracy, but provides limited insight into which geometric configurations benefit most. A breakdown by object aspect ratio, orientation, proposal quality, and angle residual would help clarify whether the gains are specifically attributable to the proposed symmetry-aware representation.

  1. GCA combines four transformed RoI representations, which may itself improve feature robustness through implicit ensembling. Comparisons with unsigned multi-frame averaging, unconstrained multi-frame fusion, and alternative parameter-matched aggregation schemes would help determine the specific advantage of enforcing target-dependent transformation laws.

4.The reported FLOPs exclude certain operations, while latency measurements primarily focus on the detection head. End-to-end inference latency, memory consumption, and throughput under different numbers of RoIs are needed to establish the practical computational overhead of SRR.

Questions for authors

see the Weaknesses

3 comments

(1/2)

Thank you for your constructive feedback highlighting the need for direct geometric and computational comparisons. Your suggestions helped us substantially strengthen the manuscript. Unless specified otherwise, AP results use final checkpoints and report mean ± sample standard deviation over three training runs.

W1 / Q — Localization and perturbations

On shared baseline RPN proposals, three-run mean refined IoU increases from 0.8504 for Baseline to 0.8597 for SRR on HRSC2016. On DIOR-R, SRR achieves 0.8026, compared with 0.7966 for Baseline. Each model uses its own trained backbone. Table 5 in §4.3 of the revised manuscript reports these comparisons.

To assess sensitivity to proposal angles, we perturb the shared positive proposals by ±5° and evaluate final checkpoints from three training runs on all 11,738 DIOR-R test images. Within each run, the three models use the same Baseline RPN proposals with initial IoU ≥0.5. Original GT assignments and each model's image feature maps remain fixed; RoI features are extracted again at the perturbed proposals. We evaluate refined IoU before NMS, averaging within GT instances and then equally across instances. The table reports mean ± sample standard deviation across runs.

Angle perturbationBaselineGCASRR
Natural0.79658 ± 0.000600.80045 ± 0.000470.80261 ± 0.00072
−5°0.78987 ± 0.000600.79511 ± 0.000630.79718 ± 0.00078
+5°0.79054 ± 0.000360.79766 ± 0.000520.79978 ± 0.00056

Relative to each model's natural condition, the mean IoU drop averaged equally over −5° and +5° is 0.00638 for Baseline, 0.00406 for GCA, and 0.00413 for SRR. GCA and SRR reduce this drop by 36.3% and 35.3%, respectively. The smaller perturbation loss is already present with GCA, while SFL improves absolute localization: SRR exceeds GCA in every run and condition. We have added the angle perturbation comparison to Appendix C.6 of the revised supplementary material.

We also examine the target transformation laws directly on 58,496 fixed pooled RoI features, comparing predictions after local axis reversals with the original predictions transformed by the target signs. Across GCA and SRR, maximum absolute differences are 3.81×10−63.81\times10^{-6} for classification logits and 1.91×10−61.91\times10^{-6} for regression outputs.

W2 / Q — Geometric configurations

For positive proposals with initial IoU ≥0.5 on DIOR-R, SRR improves refined IoU over Baseline across the groups below. Values are means over three runs. Aspect-ratio groups use GT boxes; quality groups use the common input proposals.

GT aspect ratioIoU gain over Baseline
[1,1.2)+0.0055
[1.2,2)+0.0076
[2,4)+0.0056
[4,∞)+0.0068
Initial IoUIoU gain over Baseline
[0.5,0.7)+0.00642
[0.7,0.9)+0.00445
[0.9,1]+0.00246

The following breakdowns use seed-0 final checkpoints. GT orientation uses the long-axis angle for AR≥1.2. Absolute initial coded angle residuals use le90 edge-swap targets across all aspect ratios. Each bin averages proposals within GTs, then equally across GTs.

GT orientation (°)GTsSRR − BaselineSRR − GCA
[−90,−45)12,950+0.00804+0.00266
[−45,0)10,289+0.00652+0.00277
[0,45)24,562+0.00589+0.00292
[45,90)28,973+0.00602+0.00183
Absolute coded residual (°)GTsSRR − BaselineSRR − GCA
[0,15)93,056+0.00638+0.00239
[15,30)25,865+0.00768+0.00357
[30,40)11,360+0.00869+0.00567
[40,45]6,391+0.01093+0.00801

SRR improves refined IoU over both Baseline and GCA in all four orientation intervals, and its gains increase across the four coded-residual bins. Across three runs, the mean SRR−GCA gain within 5° of the coder's 45° width–height swap boundary is 0.00691 IoU, versus 0.00216 overall, with gains in every run. We have added the full breakdowns to Appendix C.6 of the revised supplementary material.

(2/2)

W3 / Q — Target signs and multi-frame fusion

All variants use the same four frames and hidden width with SFL disabled. GCA uses fixed equal signed weights. The table reports DOTA-v1.0 results.

AggregationAP50AP75AP50:95
Unsigned mean76.86 ± 0.3050.98 ± 0.4247.73 ± 0.27
Learned fusion77.00 ± 0.1551.95 ± 0.7648.23 ± 0.14
GCA77.03 ± 0.1652.56 ± 0.6648.60 ± 0.59
Fixed nonuniform77.58 ± 0.0852.17 ± 0.5148.62 ± 0.18

Learned fusion optimizes 16 free coefficients, initialized uniformly. GCA achieves higher AP75 and AP50:95 than unsigned mean and learned fusion. On DIOR-R, mean AP50:95 is 41.93 ± 0.20 for GCA, compared with 40.70 ± 0.31 for unsigned averaging and 41.16 ± 0.25 for learned fusion with uniform initialization.

The fixed nonuniform variant retains the target signs but assigns weight 0.4 to the original frame and 0.2 to each reversed frame, with the same head parameter count as GCA. It favors AP50, whereas GCA achieves higher AP75, with similar AP50:95. GCA is our default because its equal signed weights preserve the prescribed transformation laws exactly.

W4 / Q — Latency, memory, and throughput

We measure patch-level detector latency, memory, and throughput across head architectures and RoI counts, including the FAA and Strip Head implementations compared in Table 6. The measurements cover the detector from backbone feature extraction through box decoding and NMS, including operations omitted from the FLOP counts.

RoIsHeadMedian detector latency (ms)Peak allocated memory (MiB)Throughput (patches/s)
128Baseline46.73506.821.730
128GCA47.89556.821.127
128SRR48.08580.821.004
512Baseline48.80506.820.844
512GCA50.31556.820.077
512SRR50.69580.819.910
2000Baseline54.40506.818.700
2000GCA60.77578.216.654
2000FAA70.12888.114.503
2000Strip Head73.771087.913.875
2000SRR62.40603.916.089

At the default 2,000 RoIs, SRR increases median detector latency by 8.00 ms (14.71%) and peak allocated memory by 97.1 MiB over Baseline. The latency overhead narrows to 1.89 ms at 512 RoIs and 1.35 ms at 128 RoIs. Although FAA and SRR have similar FLOPs in Table 6, SRR has 11.0% lower latency and 32.0% lower peak allocated memory. The reductions relative to Strip Head are 15.4% and 44.5%, respectively. SRR also has higher throughput than both heads. Table 6 reports mean AP75 gains of 1.61 points over FAA and 1.82 points over Strip Head, along with higher AP50:95.

We use final seed-0 checkpoints on the same 20 DOTA test patches with an RTX 2080 Ti, eager FP32 inference, and batch size 1. Latency is the median over synchronized patch calls; throughput is completed patch inferences divided by elapsed time in a separate measurement. Memory is peak GPU memory allocated by PyTorch, including model and input tensors. Data loading, preprocessing, host-to-device transfer, and whole-image splitting and merging are excluded. Table 16 in Appendix C.5 reports the comparison and measurement details.

As the discussion period is coming to a close, we wanted to kindly check whether there is anything further we can clarify or discuss. We hope our rebuttal and revisions have addressed your concerns, and we would be happy to answer any remaining questions. Thank you again for your time and feedback.

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

This paper addresses oriented object detection in two-stage detectors that extract features from rotated proposals. The key observation is that a rotated proposal’s local frame admits four equivalent axis-reversal choices, under which class and size are invariant while center offsets and angle flip sign in fixed patterns. The paper proposes Symmetry-Aware Region Representations (SRR), a detection head that makes region features follow the same transformation laws as their targets. Geometry-Consistent Aggregation (GCA) combines features from the four frames by signed averaging, making the head exactly equivariant to local axis reversals, and Symmetry-Aware Feature Learning (SFL) learns a separate linear map per feature type within that constraint. Experiments on five datasets, three detectors, and four backbones show consistent improvements in localization.

Critical strengths
  • The proposed head is exactly equivariant to local axis reversals by construction, i.e., its predictions do not depend on which of the four equivalent frames a fixed RoI feature is read in, and this holds for any learned weights.
  • Clear motivation and extensive experiments (five datasets, three detectors, four backbones, seed-averaged results, capacity-matched controls) that verify the paper’s claims.
Critical weaknesses
  • The paper presents the baseline’s failure to respect the transformation laws as the problem SRR solves, but never shows what this failure actually causes. Since the baseline always reads a proposal in one fixed frame, the lack of equivariance may not appear as any visible error.
Questions for authors
  • The paper identifies not respecting the transformation laws as a problem. Can the authors show what concrete consequence this has in the baseline, and provide evidence that SRR’s gains come from fixing it?
2 comments

Thank you for your insightful questions on the concrete role of the transformation laws, which prompted us to add direct geometric and empirical analyses clarifying their value as a structural prior. Unless specified otherwise, trained-model results use final checkpoints and report means over three runs, with ± denoting the sample standard deviation.

W1 / Q1 — Consequences of violating the transformation laws

Our motivation is to use the known target transformation laws as a structural prior for region feature learning. A head that uses one fixed frame can still make accurate predictions. However, in a constructed example approaching the periodic le90 boundary, we observe a non-vanishing jump in the baseline's decoded centers when the center sign-reversal law under uvuv is violated.

At this boundary, an angle wrap by π\pi reverses both local axes without exchanging the width and height. The pooled RoI tensors on the two sides are therefore related by UuvU_{uv} in the limit. To preserve the decoded image center, the local center offsets must reverse sign as well. Decoding the center offsets in the two frames gives

C0−Cuv=R(θ)diag⁡(w,h)diag⁡(σc)(z(x)+z(Uuvx)).C_0-C_{uv}=R(\theta)\operatorname{diag}(w,h)\operatorname{diag}(\sigma_c)\big(z(x)+z(U_{uv}x)\big).

Here, xx is the pooled RoI feature, R(θ)R(\theta) rotates from proposal axes to image axes, and w,hw,h are the proposal side lengths. The standardized center output is zz, and σc\sigma_c contains the coder's two center standard deviations. The coder uses zero center means.

We project the center outputs onto the sign-reversal law by replacing z(x)z(x) with [z(x)−z(Uuvx)]/2[z(x)-z(U_{uv}x)]/2, keeping the checkpoint and learned features fixed. We evaluate the DIOR-R baseline from seed 0, epoch 12, across all 11,738 test images. In this constructed case, legal proposals have angles −π/2+ϵ-\pi/2+\epsilon and π/2−ϵ\pi/2-\epsilon, where ϵ\epsilon is the angular distance from each side to the seam. At finite ϵ\epsilon, their orientations differ by 2ϵ2\epsilon modulo π\pi; they converge to the same unoriented rectangle as ϵ\epsilon approaches zero. As ϵ\epsilon decreases from 0.5° to 0.005° per side, the unprojected mean center jump approaches 1.19 px, while the projected jump falls from 0.1218 to 0.0013 px.

This example illustrates the geometric meaning of the center law under the double reversal. The full group also includes the two single-axis reversals and their target-specific sign patterns.

To examine whether the prescribed signs help beyond using multiple frames, we compare GCA with learned scalar fusion of the same four frames, with SFL disabled. GCA fixes the aggregation signs according to the targets, whereas learned fusion starts from uniform coefficients. Under the same DIOR-R training settings, GCA achieves 41.93 ± 0.20 AP50:95, versus 41.16 ± 0.25 for learned fusion over three runs. The learned coefficients can represent GCA's coefficients, yet GCA achieves higher mean AP. This supports using the known laws as a structural prior during training.

We also measure box refinement directly to assess physical localization beyond detection AP. Table 5 in §4.3 of the revised manuscript reports the shared-proposal comparison. Models receive identical baseline RPN proposals and GT assignments, but each extracts features with its own trained backbone. We use the final epoch-36 checkpoints for HRSC2016 and epoch-12 checkpoints for DIOR-R. Proposals have initial IoU ≥0.5. Before NMS, we average each metric over proposals within each GT and then across GTs. Angle errors use GT aspect ratios ≥1.2.

DatasetHeadIoU ↑Center ↓Size ↓Angle (°) ↓
HRSC2016Baseline0.85040.04390.05871.102
GCA0.85780.04160.05641.016
SRR0.85970.04100.05571.002
DIOR-RBaseline0.79660.06200.08693.981
GCA0.80040.05970.08543.939
SRR0.80260.05890.08453.865

Every geometric metric improves from Baseline to GCA to SRR on both datasets. Together with the aggregation comparison, these results support using the known target laws as a prior for learning accurate box refinement.

As the discussion period is coming to a close, we wanted to kindly check whether there is anything further we can clarify or discuss. We hope our rebuttal and revisions have addressed your concerns, and we would be happy to answer any remaining questions. Thank you again for your time and feedback.

Reviewer_PYa8OpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

The paper addresses region feature representation in two-stage oriented object detection by proposing SRR, which exploits the geometric symmetry associated with reversals of the local coordinate axes of proposals to extract features in multiple coordinate frames. GCA aggregates features from four coordinate frames using signed weights. SFL is further introduced to replace the shared linear layer in the GCA head with independent mappings for each feature type, while preserving equivariance to local axis reversals. The method improves the localization accuracy of oriented bounding boxes across multiple datasets.

Critical strengths
  1. The problem formulation is concrete, and the correspondence between the network architecture and the geometric targets is relatively clear.
  2. The method yields improvements across multiple combinations of detectors and backbones, with comprehensive experimental coverage.
Critical weaknesses
  1. Does modeling the local coordinate symmetry of a proposal with the full four-element group provide additional benefits over directly using the subgroup? This remains unverified. Additional experiments are recommended to establish the practical advantage of using the full four-element group.
  2. The incremental gains of SFL over GCA are relatively small on DIOR-R. Although the paper reports means and standard deviations over three training runs, more direct evidence is needed to establish the stability of these gains.
  3. Table 6 reports the latency of the detection heads for SRR and task-specific branches, but lacks measured comparisons with the baseline and GCA. I suggest adding a latency comparison of the baseline, GCA, and SRR under identical measurement conditions to quantify the actual runtime overhead introduced by each of the two modules.
  4. Although the paper provides reasonably extensive validation across multiple datasets, the current evaluation remains concentrated on remote sensing. I suggest including a dataset outside the remote sensing domain to assess the effectiveness of the design in different application scenarios.
  5. In Table 3, AP75 and AP50:95 for FAIR1M are marked with “—”. I suggest briefly explaining in the table caption why these metrics are not provided.
  6. Table 5 specifies the hidden layer widths alongside the Baseline and GCA entries. For consistency, I suggest labeling the SRR row as “SRR (1,024)” as well.
  7. The capitalization of conference and journal names is inconsistent in some references and should be standardized.
Questions for authors
  1. When using the same detector, backbone, and training settings, how much additional benefit does modeling with the full four-element group provide compared with using only the subgroup?
  2. Are the gains of SFL over GCA observed in all three training runs on DIOR-R?
  3. Does the proposed design remain effective in applications outside remote sensing?
2 comments

Thank you for your thoughtful review and constructive questions on group choice, SFL, runtime, and applicability, which helped us strengthen both our empirical validation and clarity. Unless specified otherwise, AP results use final checkpoints and report mean ± sample standard deviation over three training runs.

W1 / Q1 — Full group versus subgroups

We compare full SRR with the two-element subgroups Hu={e,u}H_u=\{e,u\}, Hv={e,v}H_v=\{e,v\}, and Huv={e,uv}H_{uv}=\{e,uv\} on DOTA-v1.0 and DIOR-R. All variants use Oriented R-CNN with ResNet-50. Each subgroup retains type-specific learning, and all variants have the same number of second-layer type-map weights.

HeadDOTA-v1.0 AP50DOTA-v1.0 AP75DOTA-v1.0 AP50:95DIOR-R AP50DIOR-R AP75DIOR-R AP50:95
Full SRR77.46 ± 0.2253.27 ± 0.2049.18 ± 0.2665.97 ± 0.4043.90 ± 0.1542.09 ± 0.16
SRR HuH_u77.73 ± 0.2052.94 ± 0.3048.98 ± 0.0765.42 ± 0.1343.87 ± 0.3341.80 ± 0.03
SRR HvH_v77.25 ± 0.5152.43 ± 0.4348.44 ± 0.2966.21 ± 0.1343.94 ± 0.1242.09 ± 0.05
SRR HuvH_{uv}77.47 ± 0.1852.59 ± 0.1648.57 ± 0.1565.97 ± 0.3543.64 ± 0.4441.87 ± 0.22

Among the two-element subgroups, the results favor HuH_u on DOTA-v1.0 and HvH_v on DIOR-R. The full group achieves competitive performance on both datasets, supporting its use as a common design choice. Each subgroup merges pairs of target transformation laws; the full group separates all four laws and learns a distinct map for each feature type. We have added these results to Table 9 of the revised manuscript.

W2 / Q2 — SFL gains in all three DIOR-R runs

Both localization metrics improve in all three runs underlying Table 4. The differences below are SRR minus GCA for each seed.

SeedΔ AP75Δ AP50:95
0+0.05+0.26
1+0.72+0.19
2+0.35+0.03

W3 — Baseline, GCA, and SRR latency

Both GCA and SFL increase measured head latency under identical conditions.

HeadMedian head latency (ms)
Baseline1.51
GCA2.91
SRR3.24

We measure these heads on the same RTX 2080 Ti in eager FP32 mode with batch size 1 and 512 RoIs, using the same 20 DOTA test patches and synchronized timing. The measurements include decomposition, reconstruction, nonlinearities, and prediction layers. On the 2080 Ti with 512 RoIs, full-detector latency is 48.80 ms for Baseline and 50.69 ms for SRR. Table 16 in Appendix C.5 reports detector costs at different RoI counts and gives the measurement details.

W4 / Q3 — Applications outside remote sensing

SRR also improves detection on LOAF-512 for fisheye pedestrians and Mothbox Training Dataset 1-1 for insects. Mean AP75 increases by 2.20 points on LOAF-512 and 2.05 points on Mothbox.

DatasetHeadAP50AP75AP50:95
LOAF-512Baseline71.67 ± 1.2137.33 ± 0.9138.43 ± 0.64
LOAF-512SRR73.37 ± 0.4939.53 ± 0.5140.10 ± 0.36
MothboxBaseline66.12 ± 0.4825.19 ± 1.5331.30 ± 0.63
MothboxSRR66.58 ± 0.3227.24 ± 0.4232.28 ± 0.24

We train Baseline and SRR separately on each dataset using Oriented R-CNN with ResNet-50 for 12 epochs and a total batch size of 8. LOAF-512 uses the released splits and COCO-style 101-point AP on the test split, with at most 100 detections per image. Mothbox retains the released image splits before forming 1,024-pixel patches. Evaluation uses area-integrated AP on test patches.

W5 — FAIR1M dashes

We have clarified in the revised Table 3 caption that dashes denote AP75 and AP50:95 metrics not returned by the official FAIR1M evaluation servers.

W6 — Hidden-width label

The SRR row in Table 6 is now labeled “SRR (1,024)”, consistently with Baseline and GCA.

W7 — Reference capitalization

Conference and journal names are now consistently capitalized throughout the bibliography.

Thank you for the response and clarifications. After considering the rebuttal, I will maintain my original score.

Comments on the paper

2 comments

Dear Reviewers,

Thank you for your time and effort in reviewing this submission. Given the mixed opinions among reviewers, I would like to kindly encourage you to engage in the discussion, consider the authors’ responses, and update your ratings as appropriate.

Thank you again for your valuable contributions!

AuthorsOpenReview ↗

Summary of Revisions

We thank all reviewers and the Area Chair for their constructive feedback. As the discussion period is closing and Reviewers yvQu and EEMT have not yet responded, we summarize our main revisions and responses below. All manuscript changes are highlighted in blue in the revised PDF.

  • Role of target transformation laws vs. multi-frame fusion (yvQu, EEMT). In our rebuttal, we show a concrete consequence of violating the transformation laws near the periodic le90 seam. The DIOR-R baseline exhibits a non-vanishing decoded center jump approaching 1.19 px, whereas enforcing the uvuv center sign-reversal law reduces the jump to 0.0013 px. In the revised paper, new Table 8 (§4.3) compares GCA against unsigned four-frame averaging and unconstrained learned fusion under the same four-frame head (SFL disabled). GCA outperforms both on DOTA-v1.0 (52.56 vs. 50.98 / 51.95 AP75, and 48.60 vs. 47.73 / 48.23 AP50:95) and DIOR-R (41.93 vs. 40.70 / 41.16 AP50:95). This shows that the gains come from the prescribed target signs rather than implicit multi-frame ensembling.
  • Direct localization accuracy and perturbation robustness (yvQu, EEMT). We moved the shared-proposal box refinement study into the main paper (Table 5, §4.3), showing monotonic gains from Baseline to GCA to SRR in refined IoU, center, size, and orientation errors on HRSC2016 and DIOR-R. New Appendix C.6 (Tables 17-19) adds breakdowns by aspect ratio, initial proposal IoU, orientation, and coded angle residual, along with ±5° proposal angle perturbations on DIOR-R. SRR improves across all geometric groups, with the largest gains on lower-quality proposals and large coded angle residuals near the 45° swap boundary (+0.01093 over Baseline, and +0.00691 for SRR over GCA across three runs). GCA and SRR also reduce Baseline's perturbation IoU drop by 36.3% and 35.3%.
  • Full group vs. subgroups and run-to-run stability (PYa8). New Table 9 (§4.3) compares the full group G\mathcal{G} against two-element subgroups (HuH_u, HvH_v, HuvH_{uv}) under a matched second-layer weight budget. Which subgroup works best depends on the dataset (HuH_u on DOTA-v1.0, HvH_v on DIOR-R), whereas the full group separates all four transformation laws and performs strongly on both. We also verified in our rebuttal that SFL improves over GCA in all three DIOR-R seeds on both AP75 and AP50:95.
  • End-to-end latency, memory, and throughput (EEMT, PYa8). New Appendix C.5 (Table 16) reports full-detector inference cost on an RTX 2080 Ti across 128, 512, and 2,000 RoIs, paired with unified head timing in Table 7 (and head-only timings of 1.51 / 2.91 / 3.24 ms for Baseline / GCA / SRR at 512 RoIs in our rebuttal). Over Baseline, SRR adds 1.35 ms at 128 RoIs, 1.89 ms at 512 RoIs, and 8.00 ms (+97.1 MiB) at 2,000 RoIs. At 2,000 RoIs, SRR has lower latency and peak memory than FAA (-11.0% latency, -32.0% memory) and Strip Head (-15.4% latency, -44.5% memory), while achieving +1.61 and +1.82 higher AP75.
  • Applicability beyond remote sensing and presentation fixes (PYa8). In our rebuttal, we evaluated SRR without modification on fisheye pedestrians (LOAF-512, +2.20 AP75, +1.67 AP50:95) and insects (Mothbox, +2.05 AP75, +0.98 AP50:95), showing that the formulation transfers beyond remote sensing. We also clarified the structural-prior motivation (§1, §3.1), explained the FAIR1M dashes in Table 3, standardized the SRR (1,024) label in Table 6, and fixed bibliography capitalization.

Reviewer PYa8 has reviewed our rebuttal and maintained their Weak Accept rating. Reviewer EEMT's points focused on diagnostic verification, fusion controls, and end-to-end profiling rather than issues with the soundness or core symmetry formulation of SRR. We have added all of these experiments, and each points to the same conclusion. Full details can be found in our responses to each reviewer, and we would welcome any further feedback before the discussion period closes.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.