acceptodds
Under review as a conference paper at ICLR 2027

Learning Voxel-wise Observation Reliability for Robust LiDAR 3D Object Detection

Abstract

Voxel-based 3D detectors achieve a favorable accuracy-efficiency trade-off, but remain vulnerable to weather- and sensor-induced corruptions. A key limitation is that existing voxel-based pipelines typically treat all occupied voxels as equally reliable, even though corruption can introduce spurious or perturbed observations with substantially different reliability. Once encoded, unreliable voxel features can be repeatedly propagated through sparse convolutions, contaminating otherwise reliable geometric representations. To address this issue, we propose VORL, a Voxel-wise Observation Reliability Learning framework for robust LiDAR 3D object detection. VORL estimates a continuous reliability score for each occupied voxel from its feature representation and lightweight observation cues, and introduces Reliability-aware Sparse Convolution (RaSC) to modulate source-voxel contributions before neighborhood aggregation. In this way, unreliable feature propagation is suppressed before feature mixing occurs, while reliable geometric evidence is largely preserved. Since voxel-wise reliability annotations are unavailable, we derive corruption-aware pseudo reliability targets from clean-to-corrupted correspondences available during corruption synthesis. Experiments on KITTI-C and nuScenes-C demonstrate consistent robustness improvements on diverse voxel-based detectors. On KITTI-C, VORL improves corrupted AP by 2.62-6.78 points across three baseline detectors while maintaining clean performance. More importantly, VORL can be seamlessly integrated into existing voxel-based detectors, while introducing only marginal computational and memory overhead.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
LiDAR 3D object detection, corruption robustness, voxel-wise observation reliability, sparse convolution, robust perception

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

This paper addresses the insufficient robustness of voxel-based LiDAR 3D object detectors under data corruption scenarios such as adverse weather or sensor failures. This paper proposes VORL, a Voxel-wise Observation Reliability Learning framework for robust LiDAR 3D object detection. VORL estimates a continuous reliability score for each occupied voxel from its feature representation and lightweight observation cues, and introduces Reliability-aware Sparse Convolution (RaSC) to modulate source-voxel contributions before neighborhood aggregation.

Critical strengths
  1. This paper is well-written.
  2. The problem of robust perception for autonomous driving in bad weather is important.
Critical weaknesses
  1. A quantitative analysis of the accuracy of the predicted voxel reliability scores is needed to demonstrate that it is correlated with the robustness improvement.
  2. For the nuScenes-C experiments, a multimodal method such as MVP should be added to the comparison.
  3. The paper looks like a stacking of modules.
Questions for authors
  1. The chosen baseline is SQD. Considering that weather and noise can also affect the image modality, how to handle the production of pseudo labels for pseudo point clouds and predict the reliability of voxels
  2. Why the metric of SQD is lower than the original ACM MM published paper?
  3. Is it suitable to add an experiment on Waymo-C?
  4. What is the research problem and novelty for this work? Because 3D perception in autonomous driving is a very ripe topic.
9 comments

(1) Reliability prediction accuracy and its impact on detection robuness. To quantitatively evaluate reliability prediction accuracy, we measure the Pearson correlation and mean absolute error (MAE) between predicted reliability scores qq and pseudo targets q∗q^* on KITTI-C using SQDNet + VORL at severity level 5, as well as the corresponding detection AP gains.

CorruptionSnowRainFogSunlightCrosstalkGaussianUniformImpulse
Pearson0.8460.8440.8980.8570.7330.6570.7060.698
MAE0.0480.0490.0980.0450.0430.1520.0960.103
AP Gain+3.32+3.17+10.29+11.91+8.01+8.07+7.21+2.26

The results show that the estimator achieves positive correlations with the pseudo reliability targets across all eight corruptions (Pearson: 0.657-0.898), as well as consistent detection improvements. Although prediction accuracy and AP gains are not directly proportional across corruption types, our additional oracle gating experiment provides further evidence of their connection to detection robustness. Replacing predicted scores with pseudo targets improves APcorAP_{\mathrm{cor}} from 67.06 to 70.93, compared with 60.28 for the baseline.

MethodSnowRainFogSunlightCrosstalkGaussianUniformImpulseAPcorAP_\mathrm{cor}
SQDNet59.7057.8061.2367.5572.8036.9952.9173.2860.28
+VORL63.0260.9771.5279.4680.8145.0660.1275.5467.06
+VORL (Oracle)67.6065.9876.6479.8781.2651.1769.2975.6370.93

Furthermore, our feature discrepancy analysis shows that VORL effectively mitigates corruption-induced feature contamination, which can otherwise degrade detection performance. All voxels are categorized into three categorizes based on pseudo reliability: Reliable-Isolated (q∗≥0.8q^* \geq 0.8, without unreliable neighbors), Reliable-Exposed (q∗≥0.8q^* \geq 0.8, with unreliable neighbors), and Unreliable (q∗≤0.3q^* \leq 0.3). Feature discrepancies are computed over spatially matched voxels.

VoxelsModelStage 1Stage 2Stage 3Stage 4
Reliable-IsolatedCorrupted0.7650.5080.5620.596
VORL0.7510.4690.5000.516
Reliable-ExposedCorrupted2.1001.4051.0530.952
VORL1.0720.8510.5860.707
UnreliableCorrupted4.5852.5791.4121.133
VORL0.8870.7250.7840.837

(2) Comparison with multimodal 3D detection method MVP. We evaluate VORL on MVP, a multimodal 3D detector that incorporates image-derived virtual points into LiDAR-based detection, using nuScenes-C at severity level 5.

MethodClean mAPClean NDSSnow mAPSnow NDSRain mAPRain NDSFog mAPFog NDSSunlight mAPSunlight NDSCrosstalk mAPCrosstalk NDSGaussian mAPGaussian NDSUniform mAPUniform NDSImpulse mAPImpulse NDS
MVP65.9969.9563.8768.2964.5268.9257.9064.3019.0934.2430.4744.0456.4160.8154.7757.3163.6067.91
MVP + VORL66.0069.9464.7769.0865.2769.4858.8465.0159.2265.0763.2168.0556.8561.2656.2758.6063.6768.00
Δ Gain+0.01-0.01+0.90+0.79+0.75+0.56+0.94+0.71+40.13+30.83+32.74+24.01+0.44+0.45+1.50+1.29+0.07+0.09

[1] MVP. Multimodal Virtual Point 3D Detection, NeurIPS 2021.

(3) Generality and unified design of VORL. We would like to emphasize that VORL is designed as a general and lightweight robustness enhancement framework for existing voxel-based 3D detectors, rather than a simple stacking of independent modules.

First, VORL demonstrates strong generality, as it can be integrated into six different voxel-based detectors across KITTI-C and nuScenes-C. Experimental results show that VORL consistently improves corruption robustness with minimal architectural modifications.

Second, VORL introduces only marginal computational and memory overhead, while keeping the pretrained detector entirely frozen, without requiring retraining or fine-tuning of the detection backbone and head.

Third, VORL does not require additional 3D box annotations under corrupted conditions.

These advantages make VORL a practical and broadly applicable framework for enhancing the robustness of existing voxel-based detectors.

(4) Reliability learning for virtual points in SQDNet. In our SQDNet experiments, we focus on robustness against weather- and sensor-induced point cloud corruptions. Specifically, we first generate virtual points offline using clean LiDAR point clouds and RGB images. We then apply synthetic corruptions to both the original LiDAR points and the pre-generated virtual points, rather than leaving the virtual points unaffected. For reliability supervision, we use the corresponding clean points and corruption metadata to construct pseudo reliability targets for both point sources following Eqs. (8)-(10). These point-level targets are then aggregated into voxel-wise reliability targets. At inference, VORL directly predicts voxel reliability from the fused voxel features and observation descriptors, without requiring clean references or corruption metadata for reliability estimation.

To further investigate robustness when RGB images are also corrupted, we conduct additional zero-shot evaluations under three settings: (a) both LiDAR and RGB are corrupted, (b) only RGB images are corrupted while LiDAR remains clean, and (c) only LiDAR points are corrupted while RGB images remain clean. For each setting, virtual points are regenerated using the corresponding RGB images and LiDAR-derived sparse depth maps. We evaluate four weather corruptions at severity level 5 using the existing SQDNet and VORL checkpoints, without retraining or fine-tuning.

(a) LiDAR + RGB corrupted

MethodSnowRainFogSunlight
SQDNet10.917.5046.8938.73
+VORL45.2039.8064.5178.03

(b) RGB Only Corrupted

MethodSnowRainFogSunlight
SQDNet80.3080.4280.4880.56
+VORL80.4280.6280.5680.57

(c) LiDAR Only Corrupted

MethodSnowRainFogSunlight
SQDNet11.568.3050.5339.54
+VORL45.9540.4766.8978.45

These results show that VORL improves average detection AP when both LiDAR and RGB are corrupted (+30.88) and when only LiDAR is corrupted (+30.46), while maintaining comparable performance under RGB-only corruption. These results suggest that VORL remains effective against point-cloud degradation even when image corruption affects virtual point generation.

(5) Clarification of SQDNet Baseline Performance. We initially attempted to use the publicly available pretrained SQDNet checkpoint but encountered compatibility issues with the current officially released code. Therefore, we retrained SQDNet using the official implementation and used the reproduced checkpoint for our robustness experiments.

It is worth noting that the original SQDNet paper reports its 3D detection results on the official KITTI test set, whereas our experiments are evaluated on the KITTI validation split used to construct KITTI-C. This difference in evaluation splits may explain the apparent discrepancy in the reported performance. Moreover, both SQDNet and its VORL variant use the same checkpoint.

(6) Evaluation on Waymo-C. We add evaluation on Waymo-C using VoxelNeXt as the baseline detector.

MethodClean L1Snow L1Rain L1Fog L1Sunlight L1Crosstalk L1Gaussian L1Uniform L1Impulse L1
VoxelNeXt74.37/71.2758.27/55.4565.30/62.3758.94/55.7239.30/37.2053.61/51.0420.82/18.3449.08/45.6266.07/61.78
+VORL74.35/71.2665.10/62.1768.27/65.3061.42/58.2168.25/65.2769.86/66.8421.61/19.3561.91/58.5967.01/63.13
MethodClean L2Snow L2Rain L2Fog L2Sunlight L2Crosstalk L2Gaussian L2Uniform L2Impulse L2
VoxelNeXt66.30/63.6650.43/48.3057.09/54.7251.19/48.6933.31/31.6746.55/44.4717.76/15.9842.29/39.5958.37/54.83
+VORL66.31/63.6356.86/54.4759.92/57.4653.39/50.8860.15/57.6461.82/59.2518.46/16.8354.11/51.3659.19/55.95

(7) Research problem and novelty of VORL. Although LiDAR-based 3D object detection has been extensively studied, robustness against corrupted observations remains a significant challenge. Existing voxel-based detectors do not explicitly model voxel-wise observation reliability, allowing unreliable features to propagate through sparse convolutions and contaminate neighboring reliable representations.

The key novelty of VORL is to explicitly learn voxel-wise observation reliability and incorporate it into sparse feature propagation. Specifically, we introduce Reliability-aware Sparse Convolution to suppress unreliable source contributions before neighborhood aggregation, together with corruption-aware pseudo supervision that requires no additional 3D box annotations under corrupted conditions. Moreover, VORL is a general and lightweight framework that can be integrated into various voxel-based detectors with minimal computational and memory overhead. These properties highlight both the methodological value and practical applicability of VORL for robust LiDAR 3D detection.

Thanks. The core issues of Q1 and Q2 have not been resolved, and I insist on the original judgment.

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

This paper focus on the problem that voxel-based 3D detectors treat all occupied voxels equally. Under adverse weather and sensor noise, corrupted or spurious measurements contaminate neighboring voxels through repeated sparse 3D convolutions, progressively degrading detection accuracy. Existing robust detection techniques operate primarily at the input, frame, or model level rather than regulating feature-level propagation.

The authors introduce VORL to explicitly model and incorporate voxel-wise observation reliability. First of all, Reliability Estimator predicts a continuous reliability score for each occupied voxel by combining initial voxel features with a 10-dimensional observation descriptor (point support, sensing range, return intensity, intra-voxel geometry, and neighborhood consistency) and local spatial context from submanifold sparse convolutions. Then Reliability-aware Sparse Convolution (RaSC) modulates active source-voxel features prior to neighborhood aggregation via a continuous, parameter-free nonlinear weighting function, suppressing corrupted evidence before feature mixing occurs while preserving reliable geometry. Because manual annotations do not exist, continuous pseudo-reliability targets are derived from clean-to-corrupted point correspondences during training-time synthetic corruption generation. Training freezes the underlying detector and optimizes only the lightweight reliability module using domain-balanced loss.

Evaluated on KITTI-C and nuScenes-C, VORL boosts corruption average precision by 2.622.62 to 6.786.78 points across baselines (Voxel-RCNN, VirConv-L, SQDNet) on KITTI-C while preserving clean detection accuracy within 0.210.21 points. On nuScenes-C, it achieves marked gains against severe noise like sensor crosstalk.

Critical strengths

(1) The authors clearly identify the voxel reliability problem in feature aggregation for corrupted point clouds. And the paper propose to train a reliability estimator and aggregate the sparse features with the predicted reliability scores.

(2) In order to train the reliability estimator, the authors use synthetic corrupted point clouds, which can be used as the target for reliability score.

(3) The experiments are extensive and performance gain demonstrates the effectiveness of the proposed VORL framework.

Critical weaknesses

(1) The authors do not provide any comparison about computational efficiency of different methods. For example, the additional reliability estimator and RaSC should consume more computation in a single forward.

(2) The three baselines are all test-time adaptation method. The reviewer is concerned about the fairness of the experiment design. Test-time adaptation method do not have access to corrupted test data and only have a model trained on clean data before testing. For VORL, it is trained on corrupted data. The authors should indicate what corrupted augmentation is used during training and how much overlap they have with the testing dataset KITTI-C and NuScenes-C.

Questions for authors

Please refer to the weakness part. The reviewer is willing to change the score if the questions are well addressed.

4 comments

(1) Computational efficiency. We thank the reviewer for this comment. We would like to clarify that Table 1 already reports inference FPS and GPU memory consumption for different detectors. To further quantify the computational overhead introduced by the reliability estimator and RaSC, we provide model complexity analysis and module-wise latency breakdown.

Model complexity analysis. We measure the number of parameters and GFLOPs for three detectors before and after integrating VORL. All experiments are conducted on a single NVIDIA RTX 4090 GPU with batch size 1. GFLOPs are averaged over 200 fixed validation samples.

MethodParams(M)GFLOPsFPSMem.(GB)
Voxel-RCNN7.58650.57938.692.05
+VORL7.59650.69837.712.16
VirConv-L13.18648.49618.832.47
+VORL13.19648.73117.702.55
SQDNet11.606189.41816.119.41
+VORL11.616189.69515.469.98

The results show that VORL introduces only 0.01M additional parameters, with GFLOPs increasing by less than 0.5% across all three detectors, while maintaining near-baseline inference efficiency.

Module-wise latency breakdown. We further analyze the inference latency of individual components using Voxel-RCNN.

MethodVoxelizationVoxel Feature EncoderReliability EstimatorSparse 3D Backbone (or RaSC)BEV backboneDetection HeadTotal
Voxel-RCNN8.785 ms0.082 ms-6.473 ms1.630 ms8.876 ms25.846 ms
+VORL8.301 ms0.088 ms1.478 ms6.692 ms1.199 ms8.760 ms26.518 ms

The reliability estimator requires only 1.478 ms, while the sparse 3D backbone latency increases by 0.219 ms after introducing RaSC. Overall, VORL increases the total inference latency by only 0.672 ms (2.6%), demonstrating that reliability-aware feature propagation introduces limited computational overhead.

(2) Fair comparisons under matched corruption exposure. We thank the reviewer for this valuable comment. We conduct two groups of experiments to address this concern. First, we fine-tuned the clean-pretrained Voxel-RCNN using the same corruption training data and configuration as VORL, optimizing the original detection loss. Although fine-tuning model achieves higher APcorAP_{\mathrm{cor}} than VORL (69.46 vs. 68.47), it significantly degrades clean APAP from 86.60 to 81.26. These results show that adapting the detector to corrupted distributions can compromise its original detection capability. In contrast, VORL does not update the pretrained detector using corrupted data; instead, it learns a lightweight reliability estimator while keeping all detector parameters frozen. Consequently, VORL improves APcorAP_{\mathrm{cor}} from 64.08 to 68.47 while largely preserving clean performance (86.40 APAP).

MethodCleanSnowRainFogSunlightCrosstalkGaussianUniformImpulseAPcorFPSMem. (GB)Peak Train Mem. (GB)Train Time
Voxel-RCNN86.6050.2045.8371.2676.0183.1331.8372.3082.0564.0838.692.05––
+ Fine-tuned81.2659.5956.5180.0380.1781.2743.0575.3879.7069.4638.422.0521.78131min
+ VORL86.4055.0950.5375.7584.6285.6936.9476.3782.7668.4737.712.160.7956min

Second, we compare PA-AUG, a robustness-oriented data augmentation method, and TripleMixer, an input purification method. VORL achieves higher APcorAP_{\mathrm{cor}} than both PA-AUG (68.47 vs. 64.81) and TripleMixer (68.47 vs. 67.71). Notably, VORL maintains near-baseline inference efficiency (37.71 FPS), whereas TripleMixer significantly reduces throughput to 3.79 FPS. VORL also requires less additional training time than both approaches (56 vs. 411 and 221 minutes, respectively).

MethodCleanSnowRainFogSunlightCrosstalkGaussianUniformImpulseAPcorFPSMem. (GB)Peak Train Mem. (GB)Train Time
Voxel-RCNN86.6050.2045.8371.2676.0183.1331.8372.3082.0564.0838.692.05––
+PA-AUG87.0649.8345.4373.4875.8484.0632.0974.7083.0864.8138.572.0821.78411min
+TripleMixer86.5258.2452.6175.2483.5784.6131.5873.7382.0867.713.793.535.43221min
+ VORL86.4055.0950.5375.7584.6285.6936.9476.3782.7668.4737.712.160.7956min

[1] PA-AUG. Part-Aware Data Augmentation for 3D Object Detection in Point Cloud, IROS 2021.

[2] TripleMixer. TripleMixer: A 3D Point Cloud Denoising Model for Adverse Weather, TIP 2025.

(3) Data overlap and training corruptions. We clarify that there is no overlap between the training and testing samples. Specifically, KITTI-C and nuScenes-C are constructed by applying synthetic corruptions to the corresponding validation sets, whereas VORL is trained using corrupted samples generated from the original training sets. Therefore, the training and evaluation samples are strictly disjoint.

Importantly, unlike conventional corruption-aware detector training, VORL does not require 3D box annotations for corrupted training samples. Instead, it only constructs voxel-wise pseudo reliability targets from clean-to-corrupted point correspondences and trains a lightweight reliability estimator, while keeping the pretrained detector frozen.

Thank you for the reply. The first concern of efficiency is well addressed. Meanwhile, in the second one, the reviewer is still concerned about the overlap problem. The reviewer understand that no exact data sample is used for training. However, if the corruption augmentation methods on VORL's training data are the similar to NuScenes-C and KITTI-C dataset, the model will learn the corruption distribution while the test-time-adaptation baselines only have access to it during testing. The reviewer is still concerned whether this is unfair for the baseline methods.

We thank the reviewer for the follow-up comment. We acknowledge that VORL has access to synthetic corruption distributions during offline training, whereas test-time adaptation (TTA) methods do not. We address this concern from three aspects.

(1) Fair comparisons under matched corruption exposure. As detailed in our previous response, we have conducted additional controlled experiments comparing VORL with Full Fine-tuning and Dual Adapter using the same clean-pretrained Voxel-RCNN and corrupted training data. Under matched corruption exposure, VORL achieves competitive corruption robustness (68.47 vs. 69.46 AP for Full Fine-tuning), while significantly better preserving clean performance (86.40 vs. 81.26 AP). It also outperforms Dual Adapter (68.47 vs. 67.58 AP) with less than half the training time (56 vs. 128 minutes).

Beyond these controlled comparisons, we also evaluate PA-AUG, a data augmentation-based method, and TripleMixer, an input purification method, to provide broader comparisons with alternative robustness strategies. These methods (PA-AUG and TripleMixer) leverage corrupted training data. VORL achieves higher corrupted AP, while requiring less training time.

(2) Clarification of different evaluation settings. We will clarify that TTA methods do not access to synthetic corruption during training in the revised manuscript.

(3) Generalization beyond training corruption conditions. To investigate whether VORL already fits the specific corruption distributions during training, we further conduct three generalization evaluations:

First, unseen severity levels. VORL is trained only at severity level 5 and consistently improves robustness at unseen severity levels 1–4 without retraining.

SeverityMethodSnowRainFogSunlightCrosstalkGaussianUniformImpulseAPcor
1SQDNet68.0267.6168.3676.9579.9880.4074.8778.8174.38
1+ VORL70.36 ± 0.3969.60 ± 0.0577.36 ± 0.0580.99 ± 0.0280.98 ± 0.0280.52 ± 0.0076.29 ± 0.0279.22 ± 0.0376.92 ± 0.05
2SQDNet66.9265.8267.7175.1878.0778.1267.7477.0872.08
2+ VORL68.60 ± 0.0566.71 ± 0.1176.85 ± 0.2380.54 ± 0.0281.06 ± 0.0478.81 ± 0.0671.92 ± 0.0678.28 ± 0.0675.35 ± 0.03
3SQDNet63.7362.8665.9172.7073.7569.8757.9776.2667.88
3+ VORL66.24 ± 0.2164.74 ± 0.1575.29 ± 0.2380.22 ± 0.0680.69 ± 0.0572.12 ± 0.0368.50 ± 0.2577.53 ± 0.0573.17 ± 0.03
4SQDNet61.6560.5663.5969.7671.5052.5454.5874.2563.55
4+ VORL64.28 ± 0.1563.04 ± 0.3774.49 ± 0.1880.09 ± 0.0781.08 ± 0.0161.10 ± 0.1564.09 ± 0.2676.70 ± 0.2070.61 ± 0.08

Second, unseen corruption categories. We evaluate weather-to-sensor and sensor-to-weather transfer, where the reliability estimator is trained exclusively on one corruption category and tested on the other.

Sensor-to-weather generalization on KITTI-C with SQDNet.

MethodSnowRainFogSunlightAvg.
SQDNet59.7057.8061.2367.5561.57
+ VORL61.73 ± 0.2359.44 ± 0.2456.57 ± 2.2179.41 ± 0.0764.29 ± 0.61

Weather-to-sensor generalizationon KITTI-C with SQDNet.

MethodCrosstalkGaussianUniformImpulseAvg.
SQDNet72.8036.9952.9173.2859.00
SQDNet + VORL80.75 ± 0.0334.27 ± 0.7152.94 ± 0.8474.41 ± 0.1160.60 ± 0.40

Third, different corruption generators. We directly evaluate VORL trained on the original KITTI-C using the Robo3D corruption benchmark without retraining. | Method | Snow | Fog | Crosstalk | Motion Blur | |---|---:|---:|---:|---:|---:| | SQDNet | 66.56 | 73.15 | 75.27 | 54.21 | | + VORL | 70.40 | 76.36 | 80.74 | 56.27 |

These results provide complementary evidence that VORL learns transferable observation-reliability patterns rather than relying exclusively on the specific corruption conditions during training. Combined with the controlled comparisons, they support the effectiveness and generalizability of VORL while acknowledging the different training access between offline robustness learning and TTA methods.

Reviewer_7Ni1OpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

The paper proposes VORL for improving voxel-based LiDAR detection under corruptions. A lightweight estimator predicts voxel reliability using voxel features, observation statistics, and neighborhood context. These scores modulate source features before sparse convolutional aggregation. The estimator is trained using corruption-dependent pseudo targets derived from synthetic clean–corrupted correspondences, while the pretrained detector remains frozen. Experiments report improvements on KITTI-C and nuScenes-C, with additional transfer studies in the appendix.

Critical strengths
  1. The problem and intervention are well motivated. Distinguishing reliable geometric observations from spurious or perturbed returns is relevant to robust LiDAR perception. Modulating individual source contributions before aggregation provides a clear mechanism for reducing their influence on neighboring features.

  2. The reported improvements span several detector architectures. On KITTI-C, VORL improves corrupted AP by 4.39, 2.62, and 6.78 points for Voxel-RCNN, VirConv-L, and SQDNet, respectively, while clean AP decreases by at most 0.21 points. The nuScenes-C results provide additional evidence that the approach is not specific to one detector.

  3. The deployment requirements are attractive. VORL preserves the detector parameters and requires neither clean references nor online optimization at inference. Table 1 reports approximately 2.5–6.0% lower FPS than the corresponding baselines, which appears reasonable given the robustness gains.

Critical weaknesses
  1. The main comparisons do not control for access to corruption-specific training. VORL receives offline training on synthetic corruptions, including point correspondences and corruption metadata. The principal competing methods—TENT, SAR, and MOS—are evaluated as test-time adaptation approaches, and the paper does not establish that they receive comparable offline corruption exposure. MOS, for example, is explicitly designed around adaptation to incoming test batches and historical checkpoints. Consequently, Tables 1–2 demonstrate a useful performance–inference-cost trade-off, but do not isolate how much improvement comes from the proposed reliability formulation versus additional corruption training. A detector trained or fine-tuned with the same corruption distribution is an important missing control. A lightweight adapter trained under a comparable budget would also help establish the value of the frozen-detector design. Training-based robustness and purification approaches discussed in Related Work are largely absent from the quantitative comparison.

  2. The meaning of “reliability” and its connection to detection quality remain insufficiently validated. Equations (8) and (10) define different semantics: the fraction of uncorrupted retained points versus an inverse displacement score. A return marked as displaced under sunlight or crosstalk receives zero point-level reliability, whereas a return with the same displacement magnitude under Gaussian noise can receive positive reliability. The paper should explain why these mechanism-dependent targets form a coherent quantity for downstream feature modulation. Table 7’s Pearson correlation of 0.783 is useful evidence that the estimator predicts the designed targets. However, correlation with these targets does not establish calibration or show that the targets measure the damage caused to detection features. Since training uses only the reliability loss, this connection is central. An experiment applying the pseudo targets directly as gates would help separate estimator errors from limitations of the target and weighting rule. Measurements of clean–corrupted feature discrepancy across layers would also provide direct evidence for the proposed feature-contamination explanation, which is currently illustrated schematically.

  3. Several details needed to reproduce and interpret the method are missing. The paper should identify the exact insertion layers for each detector and explain how reliability scores correspond to source voxels at every insertion. A stride-2 convolution can legitimately use reliability scores on its input grid, so downsampling alone does not imply an error. However, it remains unclear whether later gates reuse the original scores before the first resolution change, or require pooling or recomputation on coarser grids. It is also unclear how original observation reliability is interpreted after intervening feature mixing. Other important details include the reliability-network dimensions, corruption sampling and severity distribution, training/validation/model-selection splits, and whether frozen batch-normalization statistics remain fixed. The KITTI evaluation description should specify the difficulty setting and AP protocol. There is also an unexplained difference between the “voxel features only” results in Tables 3 and 7: 65.43 versus 65.85 AP. This could reflect retention of the context branch in Table 7, but the configuration should be stated. Similarly, Table 5 needs explicit definitions of magnitude-based targets for fake returns without clean counterparts and count-based targets for continuously perturbed points.

  4. Generalization is promising but uneven, and small differences lack uncertainty estimates. The appendix deserves credit for evaluating held-out conditions. Nevertheless, sensor-to-weather transfer reduces fog AP from 61.23 to 58.60, and weather-to-sensor transfer reduces Gaussian-noise AP from 36.99 to 35.32. Cross-dataset transfer yields a 2.53-point gain, compared with 8.72 points for in-domain training. These results suggest useful but incomplete transfer of the learned reliability function. All reported evaluations use synthetic corruptions. Evaluation with a different corruption generator or real adverse-weather observations would strengthen the claim that the estimator captures broadly transferable observation quality. Repeated runs are particularly important for the smaller ablation gains. For context, my unweighted average of Table 2’s eight corruption mAP values for TransFusion-L is approximately 35.48 for VORL and 35.28 for SAR. The efficiency advantage remains meaningful, but this accuracy difference should be interpreted cautiously without variability estimates.

Questions for authors
  1. How does VORL compare with corruption-augmented training under the same corruption exposure and a clearly reported training budget?
  2. What happens when the pseudo reliability targets are used directly for gating, and does gating measurably reduce feature corruption across layers?
  3. What are the exact layer placements, reliability-grid mappings, dataset splits, and configurations underlying Tables 3 and 7?
5 comments
AuthorsOpenReview ↗

Fair comparisons under matched corruption exposure

We thank the reviewer for this valuable comment. We agree that the original comparisons with test-time adaptation methods do not fully isolate the benefits of reliability-aware learning from additional corruption exposure. To address this concern, we conducted three groups of additional experiments using Voxel-RCNN as the common baseline.

(1) Comparison with corruption fine-tuning. We fine-tuned the clean-pretrained Voxel-RCNN using the same corruption training data and configuration as VORL, optimizing the original detection loss. Although fine-tuning model achieves higher APcorAP_{\mathrm{cor}} than VORL (69.46 vs. 68.47), it significantly degrades clean APAP from 86.60 to 81.26. These results show that adapting the detector to corrupted distributions can compromise its original detection capability. In contrast, VORL does not update the pretrained detector using corrupted data; instead, it learns a lightweight reliability estimator while keeping all detector parameters frozen. Consequently, VORL improves APcorAP_{\mathrm{cor}} from 64.08 to 68.47 while largely preserving clean performance (86.40 APAP). This demonstrates a more favorable balance between robustness to distribution shifts and retention of clean-domain performance, without modifying the pretrained detector parameters. Moreover, VORL requires considerably less training time (56 vs. 131 minutes).

MethodCleanSnowRainFogSunlightCrosstalkGaussianUniformImpulseAPcorFPSMem. (GB)Peak Train Mem. (GB)Train Time
Voxel-RCNN86.6050.2045.8371.2676.0183.1331.8372.3082.0564.0838.692.05––
+ Fine-tuned81.2659.5956.5180.0380.1781.2743.0575.3879.7069.4638.422.0521.78131min
+ VORL86.4055.0950.5375.7584.6285.6936.9476.3782.7668.4737.712.160.7956min

(2) Comparison with lightweight adapters. We introduced two residual adapters at the same backbone locations as RaSC, while freezing the original detector. The adapters were trained using the same corruption data and the original detection loss. VORL outperforms the dual-adapter baseline in APcorAP_{\mathrm{cor}} (68.47 vs. 67.58), while preserving comparable clean accuracy (86.40 vs. 86.13) and requiring substantially less training time (56 vs. 128 minutes).

MethodCleanSnowRainFogSunlightCrosstalkGaussianUniformImpulseAPcorFPSMem. (GB)Peak Train Mem. (GB)Train Time
Voxel-RCNN86.6050.2045.8371.2676.0183.1331.8372.3082.0564.0838.692.05––
+ Dual Adapter86.1352.8947.6973.5482.1483.1443.4775.2182.5667.5837.902.1021.85128min
+ VORL86.4055.0950.5375.7584.6285.6936.9476.3782.7668.4737.712.160.7956min

(3) Comparison with training-based robustness and purification methods. We further evaluated PA-AUG, a robustness-oriented data augmentation method, and TripleMixer, an input purification method. VORL achieves higher APcorAP_{\mathrm{cor}} than both PA-AUG (68.47 vs. 64.81) and TripleMixer (68.47 vs. 67.71). Notably, VORL maintains near-baseline inference efficiency (37.71 FPS), whereas TripleMixer significantly reduces throughput to 3.79 FPS. VORL also requires less additional training time than both approaches (56 vs. 411 and 221 minutes, respectively).

MethodCleanSnowRainFogSunlightCrosstalkGaussianUniformImpulseAPcorFPSMem. (GB)Peak Train Mem. (GB)Train Time
Voxel-RCNN86.6050.2045.8371.2676.0183.1331.8372.3082.0564.0838.692.05––
+PA-AUG87.0649.8345.4373.4875.8484.0632.0974.7083.0864.8138.572.0821.78411min
+TripleMixer86.5258.2452.6175.2483.5784.6131.5873.7382.0867.713.793.535.43221min
+ VORL86.4055.0950.5375.7584.6285.6936.9476.3782.7668.4737.712.160.7956min

[1] PA-AUG. Part-Aware Data Augmentation for 3D Object Detection in Point Cloud, IROS 2021.

[2] TripleMixer. TripleMixer: A 3D Point Cloud Denoising Model for Adverse Weather, TIP 2025.

AuthorsOpenReview ↗

Reliability Target Validity and Feature Contamination Analysis

We thank the reviewer for this insightful comment. We agree that the connection between pseudo reliability supervision and feature-level corruption mitigation requires further validation.

(1) Corruption-aware reliability targets. Although Eqs. (8) and (10) use different formulations, both aim to measure the usable geometric information preserved within each voxel. For sunlight and crosstalk, corruption affects a subset of returns through abnormal displacements, so we estimate reliability using the proportion of unaffected points. In contrast, Gaussian and other continuous noise may perturb nearly all points, while mildly perturbed observations can still preserve useful geometry. Therefore, we adopt a magnitude-based formulation to avoid excessively suppressing valid features. Both formulations share the same objective but account for different corruption mechanisms. Although these targets are mechanism-dependent approximations rather than perfectly calibrated reliability measures, their effectiveness is supported by Table 5 in the original manuscript.

(2) Oracle reliability gating. To directly evaluate whether the designed pseudo targets can effectively guide feature propagation, we add an oracle setting that uses the pseudo reliability targets q∗q^* for gating. As shown below, replacing predicted reliability with oracle reliability increase APcorAP_{\mathrm{cor}} from 67.06 to 70.93, compared with 60.28 for the original SQDNet. The improvement is observed across all eight corruption types. These results provide direct evidence that the designed pseudo reliability targets contain useful information for feature modulation. The additional 3.87-point improvement over VORL also suggests that reliability estimation error contribute to the remaining performance gap.

MethodSnowRainFogSunlightCrosstalkGaussianUniformImpulseAPcorAP_\mathrm{cor}
SQDNet59.7057.8061.2367.5572.8036.9952.9173.2860.28
+VORL63.0260.9771.5279.4680.8145.0660.1275.5467.06
+VORL (Oracle)67.6065.9876.6479.8781.2651.1769.2975.6370.93

(3) Clean-corrupted feature discrepancy analysis. To directly examine whether VORL mitigates unreliable feature propagation, we conduct a feature-level analysis using 3,769 KITTI validation samples across eight corruption types at severity level 5. We extract features from all four sparse backbone stages under clean, corrupted, and VORL variant model. We divide voxels into three categorizes based on pseudo reliability: Reliable-Isolated (q∗≥0.8q^* \geq 0.8, without unreliable neighbors), Reliable-Exposed (q∗≥0.8q^* \geq 0.8, with unreliable neighbors), and Unreliable (q∗≤0.3q^* \leq 0.3). For spatially matched voxels, we measure the normalized L2 discrepancy between corrupted and clean features.

VoxelsModelStage 1Stage 2Stage 3Stage 4
Reliable-IsolatedCorrupted0.7650.5080.5620.596
VORL0.7510.4690.5000.516
Reliable-ExposedCorrupted2.1001.4051.0530.952
VORL1.0720.8510.5860.707
UnreliableCorrupted4.5852.5791.4121.133
VORL0.8870.7250.7840.837

VORL reduces feature discrepancies, with particularly pronounced improvements in Reliable-Exposed and Unreliable voxels. These observations support our central motivation: reliability-aware feature propagation not only mitigates corrupted representations but also reduces their influence on neighboring reliable regions.

AuthorsOpenReview ↗

Generalization across Corruption Generatorand Evaluation Stability

(1) Zero-shot generalization across corruption generators. To evaluate generalization beyond the original KITTI-C corruption protocol, we directly transfer SQDNet + VORL, trained on KITTI-C at severity level 5, to Robo3D KITTI-C at Heavy severity, without retraining, test-time adaptation, or hyperparameter tuning. As shown below, VORL consistently improves detection performance across all three evaluated corruptions, achieving gains of +3.84 AP for Snow, +3.21 for Fog, and +5.47 for Crosstalk. These results demonstrate encouraging cross-generator transferability under unseen synthetic corruption protocols.

MethodSnowFogCrosstalk
SQDNet66.5673.1575.27
+ VORL70.4076.3680.74

[1] Robo3D: Towards Robust and Reliable 3D Perception against Corruptions, ICCV 2023.

(2) Multi-seed evaluation of generalization. We repeated the experiments in Appendix Tables 8-11 using three independent random seeds and report mean ± standard deviation.

Table 8 Cross-severitygeneralizationonKITTI-CwithSQDNet. The reliability estimator is trained only at severity level 5 and directly evaluated at lower severity levels1-4 without retraining.

SeverityMethodSnowRainFogSunlightCrosstalkGaussianUniformImpulseAPcor
1SQDNet68.0267.6168.3676.9579.9880.4074.8778.8174.38
1+ VORL70.36 ± 0.3969.60 ± 0.0577.36 ± 0.0580.99 ± 0.0280.98 ± 0.0280.52 ± 0.0076.29 ± 0.0279.22 ± 0.0376.92 ± 0.05
2SQDNet66.9265.8267.7175.1878.0778.1267.7477.0872.08
2+ VORL68.60 ± 0.0566.71 ± 0.1176.85 ± 0.2380.54 ± 0.0281.06 ± 0.0478.81 ± 0.0671.92 ± 0.0678.28 ± 0.0675.35 ± 0.03
3SQDNet63.7362.8665.9172.7073.7569.8757.9776.2667.88
3+ VORL66.24 ± 0.2164.74 ± 0.1575.29 ± 0.2380.22 ± 0.0680.69 ± 0.0572.12 ± 0.0368.50 ± 0.2577.53 ± 0.0573.17 ± 0.03
4SQDNet61.6560.5663.5969.7671.5052.5454.5874.2563.55
4+ VORL64.28 ± 0.1563.04 ± 0.3774.49 ± 0.1880.09 ± 0.0781.08 ± 0.0161.10 ± 0.1564.09 ± 0.2676.70 ± 0.2070.61 ± 0.08

Table 9 Sensor-to-weather generalization on KITTI-C with SQDNet.

MethodSnowRainFogSunlightAvg.
SQDNet59.7057.8061.2367.5561.57
+ VORL61.73 ± 0.2359.44 ± 0.2456.57 ± 2.2179.41 ± 0.0764.29 ± 0.61

Table 10 Weather-to-sensor generalizationon KITTI-C with SQDNet.

MethodCrosstalkGaussianUniformImpulseAvg.
SQDNet72.8036.9952.9173.2859.00
SQDNet + VORL80.75 ± 0.0334.27 ± 0.7152.94 ± 0.8474.41 ± 0.1160.60 ± 0.40

Table 11 Cross-dataset generalization from nuScenes-C to KITTI-C with matched CenterPoint-Voxel detector.

MethodSnowRainFogSunlightCrosstalkGaussianUniformImpulseAPcor
CenterPoint-Voxel54.6761.9866.3739.8462.1034.3260.7073.2656.65
+ VORL (In-domain)65.41 ± 0.1266.84 ± 0.1773.11 ± 0.1272.87 ± 0.9875.20 ± 0.1539.14 ± 0.5168.67 ± 0.1874.07 ± 0.0566.91 ± 0.25
+ VORL (Cross-dataset)58.46 ± 0.0663.78 ± 0.1368.44 ± 0.1046.09 ± 1.4064.64 ± 0.4435.36 ± 0.0763.27 ± 0.0773.67 ± 0.0159.21 ± 0.21

(3) Multi-seed evaluation on TransFusion-L. To assess the variability of the small accuracy difference, we conducted three independent training runs for VORL on TransFusion-L while keeping the pretrained detector frozen.

MethodSnowRainFogSunlightCrosstalkGaussianUniformImpulsemAPcor
TransFusion-L48.0954.8420.740.494.2928.0524.8156.0429.67
+ SAR50.7055.3623.907.3034.0229.4226.2655.2935.28
+ VORL52.41 ± 0.0857.18 ± 0.0623.84 ± 0.138.20 ± 1.7235.06 ± 2.7928.09 ± 0.0526.23 ± 0.1556.10 ± 0.0635.89 ± 0.41

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.