acceptodds
Under review as a conference paper at ICLR 2027

Aligning Sparse Messages for Pose-Robust Collaborative Perception

Abstract

Collaborative perception improves detection by sharing observations across vehicles and roadside units, but its benefits are limited by communication cost and pose noise. Under a limited communication budget, features useful for detection may still be difficult to match across agents. We propose SAComm, a collaborative perception framework that jointly designs sparse patch messages and receiver-side alignment and fusion. Each sender transmits a small number of compressed local feature patches, preserving local structure for both detection and matching. The receiver matches these patches against ego features and estimates a shared translation and yaw correction for each sender while accounting for matching uncertainty and geometric consistency. An adaptive gate controls how much correction is applied to the entire message, while pose reliability guides local fusion. Thus, the same sparse messages support both alignment and collaborative detection. Experiments on OPV2V, V2XSet, and DAIR-V2X-C show that SAComm maintains strong detection performance under pose noise at the lowest communication rates among the evaluated feature-level methods. On OPV2V, it achieves 86.98% BEV AP70 under strong pose noise at only 0.393 Mbps per sender.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
collaborative perception, sparse communication, pose robustness, LiDAR detection, autonomous driving

Reviews

Among:
Shape:
1234this paper, Top 15%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 3.00 over 3 reviews, above 85% of the others.
Reviewer_KcFBOpenReview ↗
Rating
2
Confidence
4
Informativeness
1
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes SAComm, combining sparse patch communication, uncertainty-aware pose correction, and reliability-guided fusion for collaborative perception. Experiments on three datasets demonstrate competitive detection accuracy under pose noise with substantially reduced communication.

Critical strengths
  1. The integration of sparse feature transmission and geometric alignment is meaningful. Table 1 demonstrates strong detection performance under pose perturbations while requiring substantially less communication than competing feature-level methods.

  2. The component ablations and pose diagnostics in Tables 2-4 provide useful evidence for the effectiveness of pose correction and local fusion, rather than relying solely on end-to-end detection results.

Critical weaknesses
  1. Table 1 compares methods at substantially different communication rates, making it difficult to attribute performance gains to the proposed architecture rather than bandwidth differences. Moreover, the methodological distinction from ERMVP and CPPC remains unclear, and CPPC is absent from the main comparison. These issues weaken both the empirical justification and the claimed novelty of SAComm.

  2. The proposed ACG module degrades performance under strong pose noise (Tables 2 and 4), while improvements on real-world DAIR-V2X-C become marginal at higher noise levels. Furthermore, evaluation relies primarily on synthetic Gaussian perturbations, with limited statistical validation. These findings raise concerns about whether the claimed robustness generalizes to realistic localization errors.

Questions for authors
  1. Can the authors provide communication-rate-matched comparisons with ERMVP, InfoCom, and Where2comm, and explain the absence of CPPC from Table 1? What measurable benefits does patch-based uncertainty modeling provide beyond ERMVP and CPPC?

  2. Why does ACG reduce AP70 from 88.12 to 86.98 and increase residual translation error from 0.369 m to 0.545 m at s=0.6s=0.6? Under what conditions does ACG provide consistent improvements?

  3. Can the authors substantiate robustness under temporally correlated or biased localization errors and report variability across random seeds or noise realizations, particularly for the marginal improvements on DAIR-V2X-C?

6 comments
AuthorsOpenReview ↗

Response to Reviewer KcFB [1/4]

Thank you for the detailed comments on communication, methodological differences, and robustness. We address the revised questions below.

W1 / Q1 (Part 1/2): Communication-rate-matched comparisons

We adjusted the spatial feature retention ratio in Where2comm, sampling and clustering ratios in ERMVP, and mask retention ratio in InfoCom to approximately match SAComm's communication rate, with model weights fixed. Reducing communication lowers zero-noise AP70 for all three baselines on all datasets, while often improving detection under strong noise. This pattern is consistent with greater reliance on ego features: fewer sender features can weaken the benefit of collaboration at zero noise while reducing interference from misaligned observations at high noise. A flatter noise curve can therefore reflect weaker collaboration rather than better use of noisy messages. At approximately matched communication rates, SAComm retains high zero-noise accuracy and achieves the highest AP70 across the tested settings. Specifically:

  • OPV2V: SAComm exceeds the strongest matched baseline, ERMVP, by 9.59 and 8.81 percentage points at s=0s=0 and s=0.6s=0.6, respectively.
  • V2XSet: The gains over the strongest matched baselines are 9.13 points at s=0s=0 (Where2comm) and 6.97 points at s=0.6s=0.6 (ERMVP).
  • DAIR-V2X-C: SAComm exceeds InfoCom by 6.28 points at s=0s=0 and performs comparably at s=0.6s=0.6, with a 0.07-point difference.

Table 1. BEV AP70 (%) ↑ at original and approximately matched communication rates. Rates are Mbps per sender at 10 Hz. O and M denote Original and Matched settings.

MethodOPV2VV2XSetDAIR-V2X-C
Mbps ↓s=0s=00.20.40.6Mbps ↓s=0s=00.20.40.6Mbps ↓s=0s=00.20.40.6
NoFusion072.9372.9472.9372.93062.4762.4762.4762.47055.4555.4755.4755.47
Where2comm O38.97689.9979.8267.1363.291.90778.8971.2463.4160.3818.32460.8658.8957.6157.04
Where2comm M0.3945675.9875.9175.9775.860.3916872.8869.1265.2363.540.5644857.5957.5857.5757.58
ERMVP O115.08187.2784.3880.1877.75115.08177.4673.9169.6267.2785.80158.8858.0056.9356.11
ERMVP M0.3936082.8281.4279.4178.170.3910470.9270.1769.0268.290.5657657.9557.3756.9756.53
InfoCom O4.94889.5982.5374.3670.724.94881.5075.3168.2565.093.71164.1860.7658.7257.74
InfoCom M0.3933678.7577.3776.0575.470.3908069.0968.5367.5067.090.5657658.8958.1357.8757.72
SAComm0.3934292.4191.2589.7686.980.3909582.0179.9877.7075.260.5658965.1761.1558.6457.79
AuthorsOpenReview ↗

Response to Reviewer KcFB [2/4]

W1 / Q1 (Part 2/2): CPPC comparison and uncertainty benefits

Method differences and contribution. ERMVP uses consensus matching, geometric verification, and error modulation on reconstructed features. CPPC optimizes a pose graph from point-cluster correspondences. SAComm uses compressed local patches to retain matching context, preserves candidate displacement distributions, and combines their means and covariances to estimate one correction per sender. Matching statistics and the applied correction are then reused to predict residual pose error and guide local fusion. Our contribution is this joint design of the sparse message and its receiver-side use for alignment and detection at very low communication rates.

CPPC comparison. Table 1 uses a common PointPillars [1] architecture to compare communication and fusion methods while limiting detector differences. CPPC was not included because it builds its sparse detector on FSD [2]. We reproduced CPPC and evaluated its full and low-rate configurations, using descriptor reduction and point sampling for the latter.

Table 2. CPPC comparison: BEV AP70 (%) ↑. Rates are Mbps per sender at 10 Hz. 8D is separately trained; 1/128 is the geometric-point sampling ratio. 128D payloads are estimated from their full fields; 8D rates use encoded byte counts.

MethodOPV2VV2XSetDAIR-V2X-C
Mbps ↓s=0s=00.20.40.6Mbps ↓s=0s=00.20.40.6Mbps ↓s=0s=00.20.40.6
CPPC 128D, full8.02291.7690.5287.4980.979.00887.8586.7583.9376.0113.00163.1362.7762.1261.74
CPPC 8D, 1/1280.57488.2887.3784.0178.840.76779.6178.8876.1671.311.45659.0858.9158.8458.87
SAComm0.39392.4191.2589.7686.980.39182.0179.9877.7075.260.56665.1761.1558.6457.79

Compared with low-rate CPPC, SAComm reduces communication by approximately 31.7%, 49.1%, and 61.1% on OPV2V, V2XSet, and DAIR-V2X-C, respectively, while achieving higher AP70 in 10 of the 12 evaluated settings. On OPV2V and V2XSet, SAComm improves AP70 at all four noise levels, with gains of 8.14 and 3.95 points at s=0.6s=0.6. On DAIR-V2X-C, SAComm improves AP70 at s=0/0.2s=0/0.2; at s=0.4/0.6s=0.4/0.6, AP70 is 0.20/1.08 points lower while communication is reduced by 61.1%.

Compared with full CPPC, SAComm uses only about 1/20–1/23 of the communication and achieves higher AP70 at every tested noise level on OPV2V. Full CPPC achieves higher AP70 on V2XSet and at nonzero noise on DAIR-V2X-C, but requires substantially more communication. These results highlight SAComm's strong detection performance at very low communication rates.

Evidence for the joint design. In Table 3 of the original submission, at approximately 0.394 Mbps and s=0.6s=0.6, individual features with pointwise fusion achieve 84.37% AP70, compared with 86.98% for 3×33\times3 patches with 3×33\times3 local fusion. Keeping the patch settings fixed but restricting fusion to 1×11\times1 reduces AP70 to 83.50%, highlighting the importance of receiver-side use of the retained context. Separately, with OPV2V model weights fixed, we replace per-match PC fitting precision matrices and residual metrics with training-subset means, retaining residual reweighting and the remaining pipeline. Table 3 below shows that this increases translation and yaw errors after PC+ACG at every tested nonzero noise level and reduces AP70 by 0.62–2.22 percentage points. Together, these results support the tested message–receiver design and the benefit of adaptive uncertainty weighting under pose noise.

Table 3. PC uncertainty weighting ablation. BEV AP70 (%); pose errors are measured after PC+ACG, with input pose as the shared reference. Both settings use the same checkpoint and evaluation inputs. Differences use unrounded values. Bold compares post-correction settings.

SettingBEV AP70 (%) ↑Translation error (m) ↓Yaw error (deg) ↓
Noise ss00.20.40.600.20.40.600.20.40.6
Input pose (reference)————0.00000.37030.74261.09330.00000.22270.45140.6969
Full model (uncertainty weighting)92.4191.2589.7686.980.02470.17860.30230.54510.02390.15600.28340.4936
PC mean replacement92.5290.6388.0684.770.02040.21440.38430.65650.01930.17930.34420.5722

References

[1] Lang et al. PointPillars: Fast Encoders for Object Detection From Point Clouds. CVPR 2019.

[2] Fan et al. Fully Sparse 3D Object Detection. NeurIPS 2022.

AuthorsOpenReview ↗

Response to Reviewer KcFB [3/4]

W2 / Q2: Role of ACG and training noise coverage

ACG suppresses unnecessary corrections to protect detection when poses are accurate or pose noise is small. Even with accurate input poses, feature differences across views can produce matching offsets, so applying the full PC correction can introduce errors. In the standard-training component ablation, adding ACG improves AP70 by 1.43 and 0.53 percentage points at s=0s=0 and s=0.2s=0.2, respectively (Table 4). Diagnostics from the same forward pass show that ACG reduces the translation error introduced by PC from 0.1787 to 0.0247 m at zero noise (Table 5), supporting its protective role in both pose correction and detection.

Under strong pose noise, an overly conservative gate can suppress useful corrections. To examine the effect of training coverage, we kept the architecture unchanged and expanded the training noise levels from s=0/0.2s=0/0.2 to s=0/0.2/0.4/0.6s=0/0.2/0.4/0.6. Both settings use 20 epochs of full-model robust fine-tuning from the same clean checkpoint. Broader training keeps 50% clean scenes and divides the remaining probability equally among the three nonzero noise levels.

Table 4. OPV2V BEV AP70 (%) and mean gate. “Without ACG” is the separately trained standard-noise ablation from the paper's Table 2.

Training / modelAP70 ↑Mean gate
Noise ss00.20.40.600.20.40.6
Standard / learned gate92.4191.2589.7686.980.1370.5340.6420.612
Broad / learned gate92.1691.0990.5188.400.1120.5780.7820.808
Standard / without ACG90.9890.7290.0988.12————

At s=0.6s=0.6, broader training raises the mean gate from 0.612 to 0.808. Raw PC translation error barely changes, while the final error falls from 0.5451 to 0.4065 m (Table 5). These results indicate that broader noise coverage helps reduce excessive suppression of useful corrections, although the final error remains higher than when applying the full PC correction.

Table 5. Mean translation error (m) ↓. PC and PC+ACG come from the same forward pass, using the raw and gated corrections, respectively. Both training settings share test-noise inputs. Bold compares PC and PC+ACG within each training/noise setting.

TrainingInputPCPC+ACG
Noise ss00.20.40.600.20.40.600.20.40.6
Standard00.37030.74261.09330.17870.18510.23030.36870.02470.17860.30230.5451
Broad00.37030.74261.09330.18760.18200.22370.36780.04720.18090.24130.4065

Together, Tables 4 and 5 show that ACG protects detection by suppressing unnecessary corrections when pose noise is small. Broader noise coverage during training reduces the suppression of useful corrections at high noise, although this trade-off remains.

Reviewer_vpP9OpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
2
Presentation
1
Soundness
2
Read review
Paper summary

SAComm addresses communication constraints and pose noise in collaborative perception by jointly designing sparse messages and receiver-side alignment and fusion. Sparse Patch Communication transmits compressed local feature patches that support both matching and detection. Observation-Based Pose Correction combines probabilistic matching with geometric consistency to estimate a shared translation and yaw correction for each sender. An Adaptive Correction Gate controls the correction magnitude, while Reliability-Guided Local Fusion accounts for pose reliability and residual misalignment when incorporating the corrected messages.

Critical strengths

The integrated design for supporting pose correction and detection with the same sparse message is logically motivated and clearly formulated mathematically.

Compressed local patches are reused for pose correction based on matching uncertainty and geometric consistency, as well as for reliability-guided fusion. The relationships between message representation, pose correction, and fusion are explicitly formulated, and the components logically reflect the intended treatment of uncertain correspondences and residual misalignment. Supporting both tasks with the same sparse message, without requiring additional alignment messages, constitutes the main contribution.

Critical weaknesses

1. Validation on recent datasets covering challenging real-world scenarios is limited.

Although the experiments cover three datasets, the evaluation focuses on vehicle detection and provides limited evidence for robustness in crowded, multi-class environments containing small objects such as pedestrians. These conditions are particularly relevant to SAComm because sparse local features may become less discriminative, increasing correspondence ambiguity. Incorrect matches could affect the shared pose correction applied to an entire sender message. The lack of evaluation under these challenging conditions therefore limits confidence in the practical robustness of the proposed matching and correction mechanisms.

2. The accuracy–cost trade-off of restricting SAComm’s communication, alignment, and fusion pipeline to a single feature scale is insufficiently explored.

SAComm constructs sparse messages from the sender feature Fi0F_i^0 and uses them for pose correction and reliability-guided local fusion to update the ego feature Fe0F_e^0. The updated feature is then combined with the original ego features Fe1F_e^1 and Fe2F_e^2 in the multiscale decoder. Although using high-resolution features is motivated by preserving fine spatial details, the paper does not compare this design with extensions that also communicate, align, and fuse features at additional scales. Consequently, the potential accuracy gains and additional communication and computational costs of a multi-scale extension remain unclear.

Questions for authors

1. Does SAComm remain robust on recent real-world benchmarks containing crowded scenes, diverse object classes, and small objects?

Could the authors provide an evaluation on a benchmark such as V2X-Real, including class-wise and density-stratified detection performance and correspondence errors? Comparing models with and without PC under otherwise matched settings would help determine whether pose correction consistently benefits detection or introduces degradation when matching becomes ambiguous, particularly in crowded pedestrian scenes.

2. Would extending SAComm’s communication, alignment, and fusion pipeline to multiple feature scales provide accuracy gains that justify the additional communication and computation?

Could the authors compare the current configuration, which constructs sparse messages from Fi0F_i^0 and updates Fe0F_e^0, with extensions that also communicate and fuse features at the additional scales? Please clarify whether pose correction would be shared across scales or estimated jointly from multi-scale correspondences. Reporting detection accuracy, communication cost, computational cost, and inference throughput would help explain the trade-off underlying the current single-scale design.

4 comments
AuthorsOpenReview ↗

Response to Reviewer vpP9 [1/2]

Thank you for the careful review and constructive suggestions. We address each point below.

W1 / Q1: Real-world, multi-class, and crowded scenes

We added detection experiments for vehicles, pedestrians, and trucks on V2X-Real-VC, including evaluation grouped by object count and a control with pose correction disabled. All methods use the same three-class evaluation protocol. As shown in Table 1, SAComm achieves 52.84% mAP30 and 41.74% mAP50 at s=0.6s=0.6 using 0.298 Mbps per sender. At this noise level, its mAP50 exceeds Where2comm and InfoCom by 9.38 and 5.14 percentage points, respectively. Enabling pose correction raises mAP50 from 39.32% to 41.74% and improves pedestrian AP50 by 2.06 percentage points.

Table 1. Full-test-set BEV AP30/AP50 (%). mAP averages the three classes. Rates include features, masks where used, and spatial indices at 10 Hz. Correction off uses the same weights.

MethodMbpss=0s=0s=0.6s=0.6
VehiclePedestrianTruckmAPVehiclePedestrianTruckmAP
Where2comm34.63477.32/75.3741.62/16.3654.27/47.8357.74/46.5269.15/55.2523.94/8.0848.72/33.7647.27/32.36
InfoCom3.76878.22/76.8942.14/18.2451.76/51.1757.37/48.7772.32/60.2324.78/8.1148.64/41.4748.58/36.60
SAComm, correction off0.29878.50/76.4843.16/19.3949.91/46.6057.19/47.4973.76/63.9830.74/12.4248.20/41.5650.90/39.32
SAComm0.29878.51/76.5043.18/19.4049.88/46.5557.19/47.4875.17/67.7734.33/14.4849.01/42.9852.84/41.74

Density-stratified results. We group frames by the number of clean cooperative GT objects: lower (≤48\leq48), medium (49–59), and higher (≥60\geq60). The groups contain 418/423/412 frames and average 21.7/36.7/66.6 pedestrian boxes per frame. At s=0.6s=0.6, correction improves pedestrian AP50 and mAP50 in all three groups (Table 2), with gains of 1.58 and 2.06 points in the highest-count group.

Table 2. Density-stratified BEV AP50 (%).

ssSettingLower countMedium countHigher count
VehiclePed.TruckmAPVehiclePed.TruckmAPVehiclePed.TruckmAP
0Correction off80.0818.3568.0655.5078.2222.4954.0551.5970.7718.4233.6440.94
0SAComm80.0618.4868.0055.5178.3322.4854.0251.6170.7818.4233.6140.94
0.6Correction off70.678.9356.5745.3962.7515.6051.7043.3557.7812.5529.9133.41
0.6SAComm73.5811.6959.2648.1867.8318.0152.8946.2561.3514.1330.9335.47

Pose errors. Table 3 compares pose errors before correction and after raw PC. At s=0.6s=0.6, raw PC reduces the mean translation error from 0.728 m to 0.455 m and the mean yaw error from 0.487° to 0.434° in the highest-count group.

Table 3. Mean pose errors at s=0.6s=0.6.

StageAllLower countMedium countHigher count
Trans. (m)Yaw (deg)Trans. (m)Yaw (deg)Trans. (m)Yaw (deg)Trans. (m)Yaw (deg)
Input0.74100.48260.75190.48850.74340.47000.72840.4873
After PC0.44980.43490.48020.45760.40880.40980.45450.4336

At s=0.6s=0.6, correspondence error—the distance between the predicted matching offset and its clean-pose reference over centers with valid supported candidates—has mean/P90 values of 0.646/1.269, 0.594/1.179, and 0.646/1.286 m in the lower/medium/higher-count groups. These detection and geometric results support useful correction in the tested crowded real scenes.

AuthorsOpenReview ↗

Response to Reviewer vpP9 [2/2]

W2 / Q2: Single-scale versus multi-scale communication

We added two-scale and three-scale extensions on OPV2V. Each transmitted scale has its own channel compression, patch selection, and local fusion. Pose correction is estimated once per sender from correspondences at F0F^0. The corrected relative pose and pose reliability are then shared across all transmitted scales.

We test approximately matched and increased communication budgets. The matched-rate extensions redistribute the original 34 patch centers across scales. The increased-budget models retain 34 centers at F0F^0 and add 17 at F1F^1; the three-scale model adds another 9 at F2F^2. As in the submitted single-scale model, the multi-scale variants use seed 3407, 30 epochs of clean pretraining, and 20 epochs of robust fine-tuning, with validation-based checkpoint selection. The single-scale reference uses the submitted checkpoint.

Table 4 shows that neither approximately matched-rate extension outperforms the single-scale model. With increased communication:

  • The two-scale model uses 49.5% more communication. AP70 improves by 0.49/0.56 percentage points at s=0/0.2s=0/0.2, but drops by 0.26/1.22 points at s=0.4/0.6s=0.4/0.6.
  • The three-scale model uses about 75.7% more communication. AP70 improves by only 0.14/0.15 points at s=0/0.2s=0/0.2, but drops by 0.43/0.88 points at s=0.4/0.6s=0.4/0.6.

In the tested configurations, additional scales provide small low-noise gains at higher communication cost, but these gains do not persist under strong noise.

Table 4. Detection performance and communication cost. BEV AP70 (%); rates are Mbps per sender at 10 Hz, counting FP32 features and a 32-bit index per center.

ConfigurationCenters K0/K1/K2K_0/K_1/K_2Mbpss=0s=00.20.40.6
Single scale F0F^0(SAComm)34/0/00.39392.4191.2589.7686.98
Two scales, matched rate17/17/00.39490.8890.2489.0986.53
Two scales, added rate34/17/00.58892.9091.8189.4985.77
Three scales, matched rate17/9/80.39491.0890.4789.0585.93
Three scales, added rate34/17/90.69192.5691.4089.3386.11

Computational cost. Table 5 reports parameter count, counted GMACs, forward time, and forward FPS. All five models use the same 20 input frames, with 10 warm-up runs per model followed by three measurements per frame. Timing includes feature extraction for both CAVs and receiver-side pose correction, feature fusion, and object detection. The single-scale F0F^0 model has the lowest measured cost and the highest forward throughput among these configurations, reaching 29.17 FPS.

Table 5. Model cost and inference throughput. RTX A6000, FP32, batch size 1, two CAVs including ego, zero pose noise.

ConfigurationParameters (M)Counted GMACsForward time (ms)Forward FPS
Single scale F0F^0(SAComm)8.131202.4934.2829.17
Two scales, matched rate8.375203.9736.4527.44
Two scales, added rate8.375203.9936.4527.43
Three scales, matched rate9.322205.4439.8325.11
Three scales, added rate9.322205.5340.0624.96

These results support the single-scale design used in this study. In the tested configurations, it achieves higher AP70 at approximately matched communication rates. Increasing the communication budget for additional scales yields small gains at zero and low noise, with higher computational cost and lower AP70 under strong pose noise.

Thank you for your detailed responses and additional experiments, and I apologize for my delayed reply.

Most of my concerns have been substantially addressed, particularly the first question, which is now supported by sufficient experimental evidence.

The multi-scale experiments also provide useful justification for the current single-scale design. However, pose correction is still estimated solely from F0F^0 and shared across scales. The potential benefits of jointly estimating pose correction from multi-scale correspondences remain unexplored. I hope the authors can address this aspect when preparing the camera-ready version.

As the concerns and limitations raised in my original review have been sufficiently addressed, I will increase my rating accordingly.

Thank you very much for carefully reading our paper and the additional experiments. We sincerely appreciate your thoughtful feedback and valuable suggestions.

We agree that jointly estimating pose correction from multi-scale correspondences is an important comparison. We overlooked this setting in our current experiments. We will add this experiment and discuss the results in the camera-ready version.

Thank you again for your time and helpful suggestions. We wish you all the best!

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

∙\bullet The paper studies collaborative perception for vehicles, where agents share information to improve 3D detection.

∙\bullet The main problem is that communication is limited, and pose noise can make shared features badly aligned.

∙\bullet The authors propose SAComm, which sends only a few compressed local feature patches instead of dense features.

∙\bullet These sparse patches are used for both detection and pose correction.

∙\bullet The receiver matches the patches with its own features, estimates a pose correction, and then fuses the corrected features using reliability scores.

∙\bullet Experiments on OPV2V, V2XSet, and DAIR-V2X-C show strong detection under pose noise with very low communication cost.

Critical strengths

∙\bullet The paper does not only compress the message, but also uses the same sparse patches for pose correction and fusion. This matters because low communication is not enough if the received features are not aligned well.

∙\bullet SAComm performs well under pose noise while using much less communication than other feature-level methods. The results are shown on three datasets, which makes the empirical evidence stronger.

To me, the motivation is significant since many existing work assumes the pose information is noise-free and sent over via the metadata. But they do not account for observation noise. This work calculates the poses from the sent information, makes the method robust.

Critical weaknesses

∙\bullet Some important design choices are not fully isolated. The method uses covariance-aware matching, adaptive gating, reliability prediction, and local fusion together. The paper has useful ablations, but it would be great to understand through a more direct ablation of the covariance/uncertainty part of pose correction.

∙\bullet Training objective has many weighted losses. The final loss combines detection, matching, pose, gate, and reliability losses with several lambda weights. This is normal for this kind of system, but the idea could be clearer about how sensitive the results are to these weights, especially since the task is detection and the overall objective is non-convex.

Questions for authors

∙\bullet How sensitive are the results to the small stability term added to the covariance, matching temperature, and uncertainty weighting used in pose correction?

∙\bullet The loss uses several lambda weights. How were these weights chosen, and were they tuned per dataset?

∙\bullet Are the lambda terms only standard loss weights, or is there any intended constrained/Lagrangian interpretation? (please consider the nonconvex loss direction for detection task while answering this)

∙\bullet It would be informative to test the pose correction under more realistic pose errors, especially for biased or temporally correlated localization errors. Please show a small set of experiments if feasible in the time constraint.

2 comments
AuthorsOpenReview ↗

Response to Reviewer BWvJ [1/2]

Thank you for the careful review and constructive suggestions. We address each point below.

W1 / Q1: Covariance/uncertainty part of pose correction

We tested sensitivity to the covariance stability scale, matching temperature, and uncertainty weighting in pose correction on OPV2V, with model weights fixed. The matching temperature was set empirically to 0.08 and kept fixed across all three datasets. Table 1 varies the stability scale σ0\sigma_0 and temperature τ\tau at inference. Halving or doubling σ0\sigma_0 changes AP70 by at most 0.23 percentage points. Temperature changes have little effect at zero noise, but doubling τ\tau reduces AP70 by 3.34 points at s=0.6s=0.6, indicating greater sensitivity to high temperatures under strong noise.

Table 1. Parameter sensitivity. BEV AP70 (%) ↑. All settings are re-evaluated using matched inputs within this table. The default results differ slightly from those reported in the paper.

Covariance Stability scale σ0\sigma_0Matching Temperature τ\tauBEV AP70 (%)
s=0s=0s=0.2s=0.2s=0.4s=0.4s=0.6s=0.6
0.05 (SAComm)0.0892.4291.0889.5487.08
0.0250.0892.2890.9389.4186.85
0.100.0892.5591.1289.6087.25
0.050.0492.5391.2589.8787.47
0.050.1692.3789.6886.8283.74

For uncertainty weighting, we replace the per-match fitting precision matrices and residual metrics in PC with fixed means from a training subset, retaining residual reweighting and the remaining pipeline. At all tested nonzero noise levels, this increases both translation and yaw errors after PC+ACG and reduces AP70 by 0.62–2.22 percentage points (Table 2). At zero noise, mean replacement slightly improves AP70 and pose errors. These results support adaptive PC weighting under the tested pose noise.

Table 2. PC uncertainty weighting ablation. Pose errors are measured after PC+ACG in both settings; input pose is the shared reference before correction.

SettingBEV AP70 (%)Translation error (m)Yaw error (deg)
Noise ss00.20.40.600.20.40.600.20.40.6
Input pose————0.00000.37030.74261.09330.00000.22270.45140.6969
SAComm (uncertainty weighting)92.4191.2589.7686.980.02470.17860.30230.54510.02390.15600.28340.4936
SAComm (PC mean replacement)92.5290.6388.0684.770.02040.21440.38430.65650.01930.17930.34420.5722
AuthorsOpenReview ↗

Response to Reviewer BWvJ [2/2]

W2 / Q2: Loss-weight selection and sensitivity

The auxiliary losses directly supervise matching, pose correction, ACG, and PR alongside detection. The loss weights were set empirically and shared across all three datasets, without an extensive hyperparameter search.

Table 3 tests sensitivity by halving or doubling one loss-weight group at a time. Each OPV2V run starts from the same clean checkpoint and uses training seed 3407 for 20 epochs of robust fine-tuning. We select the checkpoint with the lowest weighted validation loss for each run. Across all four noise levels, AP70 differs from the default model retrained under the same protocol by at most 0.41 percentage points. These results show that SAComm is relatively insensitive to the tested loss-weight changes.

Table 3. Loss-weight sensitivity: BEV AP70 (%). All five coefficients within the matching group are scaled together.

Noise ssDefault, retrainedPoseMatchingACGPR
Weight multiplier1x0.5x2x0.5x2x0.5x2x0.5x2x
092.3492.4092.4192.3592.4192.5592.4392.4392.47
0.291.0891.2590.9491.0791.2791.1091.1491.1491.15
0.489.5289.3989.4589.4589.5889.6589.6689.4689.48
0.686.8286.4186.7886.9786.5986.5186.5386.5686.78

Q3: Interpretation of the loss weights

The λ\lambda terms are fixed weights that balance detection and auxiliary supervision, not Lagrange multipliers. They were set empirically, not derived from a constrained optimization problem. For this non-convex objective, our follow-up sensitivity study (Table 3) shows stable detection AP when one loss-weight group is halved or doubled at a time.

Q4: Biased, correlated, and non-Gaussian pose errors

We evaluated the submitted OPV2V checkpoint under sequence-fixed bias plus jitter, temporally correlated AR(1) errors (coefficient 0.9), and Laplace errors, without retraining. Each noise family matches the Gaussian expected squared error separately for each horizontal translation component and yaw. Table 4 reports detection variability and pose errors before and after correction across three noise seeds (3407, 303, and 2027) at each strength.

Table 4. Detection and pose correction. AP70 is mean ± SD across noise seeds. Pose errors are averages across those seeds; “after” uses PC+ACG.

Noise familyAP70 (%)Translation before (m)Translation after (m)Yaw before (deg)Yaw after (deg)
Noise ss0.20.60.20.60.20.60.20.60.20.6
Gaussian91.13±0.0987.10±0.140.36891.10660.17890.55040.22330.67000.15850.4703
Bias + jitter91.08±0.2286.58±0.140.37841.13520.18250.56290.23360.70080.16440.4990
AR(1)91.04±0.1986.70±0.600.36951.10850.18070.55820.23310.69930.16560.5033
Laplace91.10±0.0987.09±0.130.35181.05530.17400.56390.21120.63370.15640.4672

At strength 0.6, AP70 decreases by at most 0.52 points relative to Gaussian noise, while correction reduces mean translation error by 46.6–50.4% for the three added families. Yaw error also decreases in each case. These tests show that correction remains useful under the tested bias, temporal correlation, and non-Gaussian perturbations.

Comments on the paper

1 comment
AuthorsOpenReview ↗

Updated Manuscript and Supplementary Materials

Dear Reviewers and Area Chair,

Thank you for your careful reviews and helpful suggestions. Your feedback has helped us improve our paper.

We have updated the manuscript PDF and supplementary materials based on your comments. All changes are highlighted in blue for easy reference.

As the discussion period is coming to an end, we would welcome any further questions or concerns. We would be happy to provide more details if needed. We hope you will consider these revisions in your final assessment.

Thank you again for your time and consideration. We look forward to your feedback.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.