acceptodds
Under review as a conference paper at ICLR 2027

Geometry-Conditioned Shared-Slice Fusion and Gaussian Set Refinement for Adverse-Weather 3D Detection

Abstract

Gated cameras provide range-selective observations through distinct temporal integration windows, offering complementary cues to LiDAR and RGB for 3D object detection in adverse weather. Existing multimodal fusion methods such as SAMFusion, however, commonly encode the near, middle, and far gated measurements as pseudo-RGB channels, conflating temporally distinct observations and obscuring their range-selective responses. This representation-level limitation is compounded at inference, where conventional suppression determines redundancy from confidence and geometric overlap without accounting for the physical evidence supporting each prediction. We present a representation-to-set framework that addresses both deficiencies. Geometry-Conditioned Shared-Slice Aggregation (GCSA) processes the gated measurements through independent passes of a shared visual encoder and estimates spatial slice contributions from projected depth, projection validity, and cross-view consistency. Slice-Geometry-Conditioned Gaussian Set Refinement (SG-GSR) subsequently derives candidate reliability and pairwise compatibility from gated-slice agreement, local LiDAR support, and depth consistency to modulate same-class Gaussian score decay. SG-GSR preserves box geometry and class labels and requires no additional detector training. Compared with SAMFusion, the complete framework improves condition-specific Overall 3D AP by 2.22–4.50 points across daytime, nighttime, snow, and fog, with gains across all three distance intervals.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
3D object detection, multimodal sensor fusion, gated imaging

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_TkeBOpenReview ↗
Rating
3
Confidence
4
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes GCSA and SG-GSR for adverse-weather LiDAR–RGB–gated 3D detection. GCSA preserves gated slice identity through shared independent encoding and geometry-conditioned aggregation, while SG-GSR refines same-class scores using gated, LiDAR, and depth evidence without retraining. On SeeingThroughFog, the full model reaches 69.72 Overall 3D AP40, outperforming the pseudo-RGB baseline by 3.60 points and SAMFusion by 2.22–4.50 points across day/night/snow/fog.

Critical strengths

Considering each slice separately is a great insight, and the geometric conditions are also carefully designed. I think this is very good.

1.Clear, physically grounded problem diagnosis: It identifies pseudo-RGB slice conflation and evidence-blind suppression, with a clear problem-to-method mapping.

2.Complementary, low-intrusion design: GCSA preserves slice identity via shared encoding and geometric routing; SG-GSR only refines scores without changing boxes, labels, or requiring detector retraining.

3.Solid experiments and mechanism analysis: It reports consistent gains on SeeingThroughFog (69.72 Overall 3D AP40, +3.60 over baseline), clean ablations, favorable efficiency (66.5% fewer gated parameters, 19.6% lower latency), and supplementary diagnostic checks.

Critical weaknesses

GCSA’s geometric conditions rely on projecting LiDAR into the RGB and gated views. If the sensors are not synchronized, calibration errors are large, or the projected LiDAR is extremely sparse, the geometric conditions can become unreliable, and the router’s slice weights may be misled. The paper itself acknowledges this.I think it would be worth trying to simulate point cloud misalignment to quantify how much performance degradation this causes.

Questions for authors

GCSA’s geometric conditions rely on projecting LiDAR into the RGB and gated views. If the sensors are not synchronized, calibration errors are large, or the projected LiDAR is extremely sparse, the geometric conditions can become unreliable, and the router’s slice weights may be misled. The paper itself acknowledges this.I think it would be worth trying to simulate point cloud misalignment to quantify how much performance degradation this causes. I think this paper is worth a weak accept. But I think it can be more complete if you can add data about misalignment's impact.

4 comments
AuthorsOpenReview ↗

Response to Reviewer TkeB (1/2)

Projection reliability and sparse geometry (W1 / Q1)

Reviewer concern (W1 / Q1).

GCSA’s geometric conditions rely on projecting LiDAR into the RGB and gated views. If the sensors are not synchronized, calibration errors are large, or the projected LiDAR is extremely sparse, the geometric conditions can become unreliable, and the router’s slice weights may be misled. The paper itself acknowledges this.I think it would be worth trying to simulate point cloud misalignment to quantify how much performance degradation this causes.

Thank you for the positive assessment and for suggesting a quantitative evaluation of projection reliability. To address whether unreliable geometry compromises GCSA, we distinguish perturbations of its routing conditions from perturbations of the detector's projection path. We evaluate the RGB and gated views separately on all 1,941 test frames, using fixed checkpoints without retraining or additional Gaussian Soft-NMS/SG-GSR.

Controlled perturbations. For each view, we test yaw offsets of ±0.1°, ±0.25°, ±0.5°, and ±1.0°, and lateral translations of ±2, ±5, ±10, and ±20 cm, separately: 16 signed settings. Cached view-specific depth geometry is backprojected, transformed, and reprojected. Router-only perturbations change GCSA's geometric condition, including validity and cross-view discrepancy. Projection-path perturbations additionally change the corresponding extrinsic and projection matrices used by the detector. Images, the other view's geometry, and raw LiDAR detector points remain unchanged.

Sensitivity to unreliable routing conditions. Across all 16 settings, the maximum AP drop is 0.1147 for gated geometry and 0.0588 for RGB geometry (Table 1A). These results directly address the concern about corrupted geometric conditions: detection accuracy exhibits limited sensitivity when the perturbation is confined to GCSA's routing inputs within the tested ranges.

Comparison under projection-path perturbations. Table 1B compares all three methods under the same 16 settings for each view. Perturbing the detector's projection path has a larger effect than perturbing routing conditions alone, particularly for RGB. Nevertheless, GCSA achieves higher absolute AP than both baselines at every tested setting. Minimum margins over uniform aggregation and pseudo-RGB are 1.5918/2.4898 AP for gated perturbations and 1.2885/2.8058 AP for RGB perturbations. This establishes a retained accuracy advantage under the tested projection errors; the comparison distinguishes this advantage from the magnitude of each model's degradation.

Table 1. Overall 3D AP₄₀ drops under controlled perturbations. Each mean includes all 16 signed settings with equal weight; each maximum covers those same settings. Drops subtract perturbed AP from the corresponding model's own unperturbed AP; negative values denote improvements. Uniform denotes shared-slice encoding with uniform aggregation.

Perturbed viewModelMean AP dropMaximum AP drop
A. Routing geometry only
GatedGCSA0.04660.1147
RGBGCSA0.02350.0588
B. Complete projection path
GatedPseudo-RGB-0.00170.6156
GatedUniform-0.00520.3910
GatedGCSA0.17620.6293
RGBPseudo-RGB1.76208.8451
RGBUniform1.52508.4808
RGBGCSA1.60828.8962
AuthorsOpenReview ↗

Response to Reviewer TkeB (2/2)

Projection-error magnitude and sparse geometry

Effect of perturbation magnitude. Table 2 reports all 16 signed RGB projection-path settings. GCSA's AP drop is 0.1728–0.3320 at ±0.1° yaw and reaches 8.8962 at -1°, the largest tested angular magnitude. At that same -1° setting, pseudo-RGB and uniform aggregation drop by 8.4707 and 8.4808 AP, respectively, while GCSA retains the highest AP (60.3861). Thus, the large loss occurs under a specified projection error and is also present in the baselines; it should be interpreted alongside the matched baseline response and the smaller routing-only effects in Table 1A.

Table 2. Overall 3D AP₄₀ for every tested RGB projection-path perturbation, with both signs retained at each magnitude. The unperturbed row gives this study's own reference for each model. Each row uses the same perturbation across models; yaw and lateral translations are applied separately.

RGB perturbationPseudo-RGBUniformGCSA
Unperturbed66.051067.171869.2823
Yaw -0.1°65.873466.867269.1095
Yaw +0.1°65.846867.296368.9503
Yaw -0.25°65.267266.006368.4528
Yaw +0.25°65.205266.959168.3469
Yaw -0.5°63.376664.200866.2247
Yaw +0.5°62.336865.266466.5549
Yaw -1°57.580358.691060.3861
Yaw +1°57.205960.043362.6721
Lateral -2 cm66.061067.042369.1833
Lateral +2 cm66.087967.201569.0639
Lateral -5 cm66.075366.874469.2125
Lateral +5 cm65.966867.388269.1781
Lateral -10 cm65.832466.561169.1120
Lateral +10 cm65.737567.283369.0938
Lateral -20 cm65.119365.913068.3662
Lateral +20 cm65.052066.753868.8783

Sparse projected geometry. We retain 50%, 25%, or 10% of valid native gated-depth samples before resizing, using identical nested random masks across models. GCSA retains the highest AP at every tested level (Table 3). Its largest drop is 0.1187 AP, at 25% retention. With only 10% retained, it achieves 69.2486 AP, a drop of 0.0658 AP from its unperturbed reference. This supports tolerance to sparse projected geometric conditions under the tested protocol.

Table 3. Overall 3D AP₄₀ under gated projected-depth thinning. The 100% row gives this study's own unperturbed references. Raw LiDAR detector points are retained.

Native samples retainedPseudo-RGBUniformGCSA
100%66.051767.101669.3144
50%66.103967.140669.2679
25%66.125967.100169.1957
10%66.092167.156469.2486

Together, these results quantify limited sensitivity to corrupted routing conditions and a retained accuracy advantage under the tested projection errors and projected-depth sparsity.

Reviewer_TkeBOpenReview ↗

Your response is very good and addresses my concern. I am willing to maintain a positive score of 3, but I do not think this paper deserves a 4.

Your response is very good and addresses my concern. I am willing to maintain a positive score of 3, but I do not think this paper deserves a 4.

Thank you for your positive assessment and for confirming that our response addresses your concern. We appreciate your constructive suggestion to evaluate geometric reliability. The resulting perturbation and sparsity analyses have strengthened our experimental evaluation and are included in the revised manuscript.

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

The motivation is reasonable, the method design is coherent, but the quality of the paper is low / There is insufficient correlation with adverse weather conditions / The validity of the module is questionable.

The paper proposes a representation-to-set framework with physical motivation. GCSA handles the utilization of gated evidence in the representation stage, while SG-GSR deals with it in the prediction set stage. The methods have certain practical value. However, the current writing and illustration affect readability. The related work is not comprehensive enough. Severe weather is mainly used as an experimental scenario rather than a method constraint. The effectiveness of SG-GSR seems questionable as well.

Critical strengths
  1. The motivation is reasonable. The paper focuses on the possibility that gated slices may be prematurely mixed or ignored during the feature fusion and candidate box selection stages, and the problem has a clear physical basis.

  2. The method and motivation are basically consistent. GCSA corresponds to slice preservation in the representation stage, SG-GSR corresponds to sensor evidence refinement in the prediction set stage, and the overall logic is coherent.

  3. The method has certain practicality. The shared encoder controls the number of parameters, and SG-GSR does not require re-training of the detector, making it convenient to integrate into the existing detection process.

  4. The ablation experiments are relatively complete. The paper distinguishes the roles of pseudo-RGB, slice preservation, geometry routing, and Gaussian refinement, which can support the main design choices.

Critical weaknesses

Writing and drawing need improvement. The motivation diagram did not sufficiently enhance the intuitiveness of the problem, and the block diagram had too many formulas, which made it less readable. The methods section also emphasized the definition of formulas more, lacking explanations of "why do it, why do it this way, and what are the benefits of doing so". For methods mainly focused on system design, the intuitive process and design purpose should be further emphasized.

The correlation between the method and adverse weather is weak. The topic of the paper is three-dimensional detection of adverse weather, but the method does not address the limitations and challenges corresponding to adverse weather. Adverse weather seems more like an important application and verification scenario.

The effectiveness of SG-GSR still needs to be strengthened. Although performance is not the most important factor, the results presented still made me question the significance of SG-GSR. In Table 3, SG-GSR improved by approximately 0.5 AP compared to GCSA, and in Table 5, it only improved by approximately 0.06 AP compared to the ordinary Gaussian decay, with a small difference, which might be close to performance fluctuations.

The coverage of related work is not comprehensive. The paper mainly focuses on SAMFusion, but lacks a summary of the recent widely studied work on three-dimensional detection of adverse weather using LiDAR-radar/4D radar.

Questions for authors
  1. Explain the overall optimization strategy and model of the paper

  2. The response method is more like a general multimodal fusion method for gated cameras rather than a method for three-dimensional detection in adverse weather conditions

  3. Positive response to the slight gain of SG-GSR, distinguishing the overall benefit of Gaussian decay from the additional benefit of sensor-conditioned evidence.

  4. Complement recent work on adverse weather fusion using LiDAR-radar/4D radar, explaining the relationship and differences between this paper and these works.

4 comments
AuthorsOpenReview ↗

Response to Reviewer XBRC (1/4)

1. Presentation, motivation, and optimization (W1 / Q1)

Reviewer comment (W1).

Writing and drawing need improvement. The motivation diagram did not sufficiently enhance the intuitiveness of the problem, and the block diagram had too many formulas, which made it less readable. The methods section also emphasized the definition of formulas more, lacking explanations of "why do it, why do it this way, and what are the benefits of doing so". For methods mainly focused on system design, the intuitive process and design purpose should be further emphasized.

Reviewer question (Q1).

Explain the overall optimization strategy and model of the paper

Thank you for highlighting the need for clearer motivation and presentation. We have revised the Introduction, Figures 1–2, and Sections 3.1–3.3 to make the design rationale, processing flow, and roles of the two modules more explicit.

Motivation and problem formulation. The revised Introduction and Figure 1 clarify two limitations: premature mixing of range-selective gated slices during feature extraction, and suppression decisions that do not explicitly account for candidate-specific sensor support. These limitations motivate GCSA before cross-modal interaction and SG-GSR after candidate generation, respectively. The text now explains how preserving slice identity supports spatially adaptive aggregation, and how candidate-aligned evidence informs competition among overlapping predictions.

Clearer figures and methodological exposition. Figure 2 now presents SG-GSR through three explicit steps: extracting candidate evidence, conditioning pairwise decay, and refining candidate scores. Detailed intermediate expressions have been replaced with descriptive operations, while the core score-update equation is retained. Sections 3.1–3.3 have also been reorganized to explain the purpose of each operation before its mathematical definition. We consolidate the displayed equation blocks from eight to six and express auxiliary definitions inline, while maintaining consistent notation between the figure and text.

Overall model and optimization. Training proceeds progressively from LiDAR-only to LiDAR–RGB and then LiDAR–RGB–gated detection. During the final stage, the LiDAR and RGB feature extractors are frozen, while the gated encoder, GCSA routers, cross-modal interaction layers, decoder, and detection heads are jointly optimized using the detector's heatmap, classification, and box-regression objectives. SG-GSR is subsequently applied to the trained detector outputs without additional optimization, modifying only confidence scores while preserving box geometry and class labels. Architecture, initialization, and optimization details are provided in Section 4.1 and Supplementary Sections A.2–A.4.

AuthorsOpenReview ↗

Response to Reviewer XBRC (2/4)

2. Connection to adverse-weather sensing (W2 / Q2)

Reviewer comment (W2).

The correlation between the method and adverse weather is weak. The topic of the paper is three-dimensional detection of adverse weather, but the method does not address the limitations and challenges corresponding to adverse weather. Adverse weather seems more like an important application and verification scenario.

Reviewer question (Q2).

The response method is more like a general multimodal fusion method for gated cameras rather than a method for three-dimensional detection in adverse weather conditions

Connection to adverse-weather sensing. Our design is motivated by the spatially and range-dependent degradation of sensor evidence under adverse weather. Gated imaging provides complementary, range-selective observations through temporal integration. GCSA preserves this structure by independently encoding the slices and jointly using slice appearance and projected geometry to determine their spatial contributions. This allows the representation to retain and selectively aggregate gated evidence when visual and geometric observations are degraded. At the prediction stage, SG-GSR further incorporates candidate-aligned gated and geometric evidence into score suppression, accounting for differences in sensor support among overlapping predictions. Together, these operations connect the sensing characteristics of adverse weather to both feature aggregation and prediction-set refinement.

Weather-specific evaluation. We compare the LiDAR–RGB baseline from Table 3 of the manuscript with the complete model on the same evaluation frames, reporting results separately for clear, snowy, and foggy conditions. We additionally examine pedestrians, whose limited spatial extent and sparse LiDAR support make detection particularly challenging under degraded sensing. The complete model improves Overall AP by 2.2298 in snow and 3.0861 in fog. The corresponding pedestrian gains are 5.2809 and 6.7003 AP, compared with 0.7410 AP in clear weather. For both overall and pedestrian performance, the 95% paired sequence-level bootstrap intervals remain above zero in snow and fog. These results support the effectiveness of the complete framework under these adverse conditions, with particularly substantial benefits for pedestrian detection.

Table 1. Overall 3D AP₄₀ under different weather conditions. Δ denotes the improvement over the LiDAR–RGB baseline from Table 3 of the manuscript.

WeatherLiDAR–RGBFull modelΔ AP
Clear63.786665.9574+2.1708
Snow63.274665.5044+2.2298
Fog52.135855.2219+3.0861

Table 2. Pedestrian 3D AP₄₀ averaged over the three distance ranges.

WeatherLiDAR–RGBFull modelΔ AP
Clear47.442748.1837+0.7410
Snow62.021667.3025+5.2809
Fog56.932663.6329+6.7003
AuthorsOpenReview ↗

Response to Reviewer XBRC (3/4)

3. Effectiveness of SG-GSR (W3 / Q3)

Reviewer comment (W3).

The effectiveness of SG-GSR still needs to be strengthened. Although performance is not the most important factor, the results presented still made me question the significance of SG-GSR. In Table 3, SG-GSR improved by approximately 0.5 AP compared to GCSA, and in Table 5, it only improved by approximately 0.06 AP compared to the ordinary Gaussian decay, with a small difference, which might be close to performance fluctuations.

Reviewer question (Q3).

Positive response to the slight gain of SG-GSR, distinguishing the overall benefit of Gaussian decay from the additional benefit of sensor-conditioned evidence.

Effectiveness of sensor-conditioned refinement. The +0.4944 AP improvement over the GCSA detector comprises +0.4339 AP from ordinary Gaussian score decay and an additional +0.0605 AP from sensor conditioning. To isolate the latter contribution, Gaussian Soft-NMS and SG-GSR are evaluated on identical frozen GCSA candidates, preserving box geometry and class labels.

To assess the stability of this incremental improvement, we perform 2,000 paired bootstrap resamples over 484 acquisition sequences. The resulting 95% confidence interval for the Overall AP gain is [0.0365,0.0793], supporting a positive contribution under sequence resampling. We further examine weather-specific and pedestrian-specific performance to identify the conditions and categories that benefit from sensor-conditioned refinement.

Pedestrian detection poses particular challenges under adverse weather, as limited spatial extent and sparse LiDAR returns compound the degradation of visual and geometric cues, especially at medium and long ranges. Pedestrian-specific analysis therefore provides a targeted assessment of this sensor-conditioned suppression under degraded sensing. Accordingly, Tables 3 and 4 report overall and pedestrian-specific performance across clear, snowy, and foggy conditions, using identical GCSA candidates and fixed post-processing configurations.

The results show larger observed gains in snow and fog, with particularly pronounced improvements for pedestrians. Pedestrian gains reach +0.4026 AP at 50–80 m in snow and +0.4844 AP at 30–50 m in fog. The 95% bootstrap intervals remain above zero for both overall and mean pedestrian gains across all three weather groups. On the full test set, pedestrians account for approximately 95.8% of the aggregate AP improvement. This concentration of benefits in the pedestrian category explains why averaging across all three categories yields a smaller overall increment.

Table 3. Overall 3D AP₄₀ under different weather conditions. Δ denotes the improvement over matched Gaussian Soft-NMS.

WeatherGaussianSG-GSRΔ AP95% bootstrap CI
Clear65.920165.9574+0.0372[0.0115,0.0664]
Snow65.419765.5044+0.0848[0.0595,0.1225]
Fog55.112655.2219+0.1092[0.0247,0.1743]

Table 4. Pedestrian 3D AP₄₀ averaged over the three distance ranges. Both tables use 2,000 paired sequence-level bootstrap resamples of the fixed predictions; CI denotes confidence interval.

WeatherGaussianSG-GSRΔ AP95% bootstrap CI
Clear48.075448.1837+0.1083[0.0321,0.1944]
Snow67.060467.3025+0.2421[0.1684,0.3565]
Fog63.321563.6329+0.3114[0.0633,0.5114]
AuthorsOpenReview ↗

Response to Reviewer XBRC (4/4)

4. Related work on LiDAR–4D radar fusion (W4 / Q4)

Reviewer comment (W4).

The coverage of related work is not comprehensive. The paper mainly focuses on SAMFusion, but lacks a summary of the recent widely studied work on three-dimensional detection of adverse weather using LiDAR-radar/4D radar.

Reviewer question (Q4).

Complement recent work on adverse weather fusion using LiDAR-radar/4D radar, explaining the relationship and differences between this paper and these works.

Thank you for pointing out this omission. We have expanded the “Multi-modal 3D detection” paragraph in Section 2 to discuss recent LiDAR–4D radar fusion methods for adverse-weather detection, including L4DR (Huang et al., AAAI 2025) [1] and DLR-Fusion (Chae et al., ICCV 2025) [2].

Relationship to prior work. L4DR combines foreground-aware denoising with multi-scale gated fusion to address heterogeneous sensor degradation. DLR-Fusion separately encodes radar power and Doppler information and integrates them with LiDAR through multi-path iterative interaction. These methods share our objective of exploiting complementary, modality-specific evidence for robust detection under adverse weather.

Distinct focus of our approach. Our method investigates the range-selective structure of gated NIR observations, whose temporal integration windows produce distinct distance-dependent responses. GCSA preserves slice identity through independent passes of a shared encoder and determines spatial aggregation weights jointly from slice appearance and projected geometry. SG-GSR subsequently uses candidate-aligned gated and geometric evidence to condition score suppression without additional training. Our work therefore complements radar-based fusion by examining how gated-slice evidence can be preserved during representation learning and incorporated into prediction-set refinement. This distinction is now stated explicitly in the revised related-work discussion.

References

[1] Xun Huang et al. L4DR: LiDAR-4DRadar Fusion for Weather-Robust 3D Object Detection. AAAI, 2025. doi:10.1609/aaai.v39i4.32397.

[2] Yujeong Chae, Heejun Park, Hyeonseong Kim, and Kuk-Jin Yoon. Doppler-Aware LiDAR-RADAR Fusion for Weather-Robust 3D Detection. ICCV, 2025. CVF Open Access.

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

The paper addresses LiDAR–RGB–gated-NIR 3D detection in adverse weather on SeeingThroughFog and proposes two additions to a DeepInteraction-based detector.

  1. GCSA. The near/middle/far gated slices are encoded separately by a shared ResNet-50-FPN instead of being packed as pseudo-RGB. A per-level router, conditioned on the slice features and on projected-LiDAR depth/validity/discrepancy maps, predicts per-pixel softmax weights to aggregate the slices.
  2. SG-GSR. A training-free extension of Gaussian Soft-NMS. The pairwise decay is modulated by gated-slice agreement, depth compatibility and a per-candidate reliability score.

Evaluation uses a reconstructed split, with all baselines retrained. Results include day/night/snow/fog comparisons against baselines, a progressive modality ablation, component ablations, a gated-encoder efficiency study and supplementary statistical audits of SG-GSR. The full model reaches 69.72 Overall 3D AP40, +3.60 over the pseudo-RGB tri-modal model, and is reported +2.22–4.50 above SAMFusion.

Critical strengths

S1. Controlled evidence that preserving gated-slice identity helps. Under identical initialization and training (Tables 3–4):

  • Shared per-slice encoding with plain uniform averaging improves Overall AP from 66.12 to 67.14.
  • Geometry-conditioned routing reaches 69.22, with gains of +1.78 to +4.16 in every range bin.
  • This is close to three independent encoders (69.31) at about one third of the gated-branch parameters (Fig. 4). Note that this matches parameters, not computation; see W1(a).
Critical weaknesses

W1. Whether the reported gains can be attributed to the proposed design.

(a) Compute-matched comparison for GCSA. The +3.1 AP of GCSA over pseudo-RGB is the paper's central evidence, but the two settings use very different amounts of computation.

  • Pseudo-RGB runs the gated backbone once. GCSA runs the shared ResNet-50-FPN three times, once per slice, so the gated-branch computation is roughly 3× higher.
  • Gated-branch latency in Fig. 4 is 22.7 ms for GCSA vs. 9.9 ms for pseudo-RGB. The parameter counts are matched (27.1M vs. 26.9M), but computation is not.
  • Within Table 4, all slice-preserving variants share the same three-pass encoder. The +2.08 AP from routing (uniform averaging 67.14 → full GCSA 69.22) is therefore compute-matched.
  • The step that the paper's motivation relies on is pseudo-RGB → uniform averaging (+1.02 AP), framed as the value of preserving slice identity. That step coincides with the ~3× increase in computation.

It is therefore unclear how much of the slice-preservation benefit comes from keeping the slices separate, and how much simply comes from additional backbone computation. I would like to see a compute-matched pseudo-RGB baseline, for example a deeper or wider backbone or a higher input resolution with comparable FLOPs and latency. FLOPs and end-to-end latency for all variants would also help. If GCSA still clearly outperforms such a baseline, the paper's central claim would be stronger.

W2. The methodological contribution seems modest relative to how it is framed, especially for SG-GSR.

  • GCSA builds on established ingredients: a weight-shared multi-branch encoder and per-pixel softmax gating over branches. These relate to reliability-weighted fusion such as the entropy-steered fusion of Seeing Through Fog (Bijelic et al., 2020).
  • SG-GSR is presented as one of three contributions. In Table 5, its sensor conditioning adds +0.06 Overall AP over Gaussian decay (69.66 → 69.72), and the reliability term adds 0.0005.
    • Given the additional hyperparameters and the class-specific configuration, the authors may want to discuss whether this gain justifies the added complexity.
    • A comparison with existing adaptive or learned suppression methods would help position SG-GSR: Adaptive NMS (Liu et al., CVPR 2019), FeatureNMS (Salscheider, 2020), learned NMS / joint rescoring (Hosang et al., CVPR 2017).
Questions for authors
  1. Compute-matched GCSA comparison.
    • Could the authors add a pseudo-RGB baseline with gated-branch FLOPs and latency comparable to GCSA, for example a deeper or wider backbone or a higher input resolution?
    • Could the authors report FLOPs and end-to-end latency for the pseudo-RGB, uniform-averaging, GCSA and independent-encoder variants?
    • This would separate the benefit of slice preservation from the ~3× gated-branch computation.
  2. Matched post-processing.
    • Could the authors report SAMFusion and DeepInteraction in the format of Table 3, with and without Gaussian Soft-NMS?
5 comments
AuthorsOpenReview ↗

Response to Reviewer uKhd (1/4)

Thank you for recognizing the gated-slice motivation and for identifying the computation and contribution-attribution questions. We address the overlapping concerns in three parts below; reviewer quotations are excerpts from the original review.

1. Compute attribution and efficiency (W1 / Q1)

Reviewer comments (W1 / Q1; excerpts).

Pseudo-RGB runs the gated backbone once. GCSA runs the shared ResNet-50-FPN three times, once per slice, so the gated-branch computation is roughly 3× higher.

Could the authors add a pseudo-RGB baseline with gated-branch FLOPs and latency comparable to GCSA, for example a deeper or wider backbone or a higher input resolution?

Could the authors report FLOPs and end-to-end latency for the pseudo-RGB, uniform-averaging, GCSA and independent-encoder variants?

Response. We agree that shared parameters do not equal matched computation: the original pseudo-RGB-to-uniform comparison changes both slice representation and encoder passes. To address this confound, we evaluated the suggested higher-resolution pseudo-RGB control, selecting its resolution by proximity to GCSA's branch MAC count before accuracy evaluation.

Table 1. Gated encoder and aggregation/adapter (RTX 3090, FP32, batch one). Pseudo-RGB/GCSA latencies follow Figure 4; the higher-resolution latency comes from the added control experiment. Accuracy is detector 3D AP₄₀, before additional Gaussian or SG-GSR.

Gated variantInput sizeGMACLatency (ms)Overall AP
Pseudo-RGB448×80049.6609.87766.1196
Higher-resolution pseudo-RGB768×1408149.86221.84867.5484
GCSA448×800150.46622.71769.2249

At a branch MAC difference of approximately 0.40%, GCSA exceeds the higher-resolution control by 1.6765 AP. Within the shared three-pass architecture, uniform aggregation obtains 67.1423 and GCSA obtains 69.2249, a further 2.0826 AP. These comparisons support retaining range-selective slice features for geometry-conditioned aggregation. The pseudo-RGB-to-uniform gain alone cannot isolate slice preservation from extra computation.

Complete-detector measurements. We also measured all four requested architectures with a common implementation and timing protocol.

Table 2. GFLOPs use 1 MAC=2 FLOPs for counted operations. Detector latency includes transfer, preprocessing, voxelization, fusion, decoding and CPU output; the last column additionally includes serialized warm-cache loading and preprocessing.

VariantParams (M)GFLOPsDetector (ms)With I/O (ms)
Pseudo-RGB84.736528.480160.208243.477
Uniform aggregation84.736728.347168.117268.210
GCSA84.986731.279176.223266.056
Independent encoders138.690731.302180.325268.626

The benchmark uses RTX 3090, batch one and FP32 tensors with TF32 enabled, with 96 timed samples and eight samples for operation counts. Offline depth-map generation and additional score refinement are excluded. Normalization, softmax, interpolation, indexing and other uncounted operations are excluded from the MAC convention; timings are measured separately. The independent encoder is profiled for architecture cost, without newly validating its historical paper AP. The host was shared with other jobs, so the I/O-inclusive timings are not an isolated deployment benchmark.

AuthorsOpenReview ↗

Response to Reviewer uKhd (2/4)

2. Method contribution and SG-GSR (W2)

Reviewer comments (W2; excerpts).

GCSA builds on established ingredients: a weight-shared multi-branch encoder and per-pixel softmax gating over branches. These relate to reliability-weighted fusion such as the entropy-steered fusion of Seeing Through Fog (Bijelic et al., 2020).

SG-GSR is presented as one of three contributions. In Table 5, its sensor conditioning adds +0.06 Overall AP over Gaussian decay (69.66 → 69.72), and the reliability term adds 0.0005.

Given the additional hyperparameters and the class-specific configuration, the authors may want to discuss whether this gain justifies the added complexity.

Response. Shared encoding and softmax gating are established ingredients. GCSA's specific design preserves near/middle/far identities within the gated modality, then aggregates their features using slice appearance and projected depth, validity and cross-view discrepancy. This makes the range-selective structure of the observations available at aggregation. The uniform-to-GCSA comparison in Section 1 provides the direct evidence for this geometry-conditioned aggregation design. We discuss our design in relation to prior work on reliability-weighted multimodal fusion [1].

SG-GSR under adverse weather. We acknowledge that SG-GSR provides a modest aggregate improvement over matched Gaussian Soft-NMS (+0.0605 AP). Its methodological contribution lies in conditioning Gaussian suppression on candidate-specific gated-slice and geometric evidence. This enables suppression strength to reflect the sensor support of competing predictions while preserving their geometry and class labels.

Pedestrian detection poses particular challenges under adverse weather, as limited spatial extent and sparse LiDAR returns compound the degradation of visual and geometric cues, especially at medium and long ranges. Pedestrian-specific analysis therefore provides a targeted assessment of this sensor-conditioned suppression under degraded sensing. Accordingly, Tables 3 and 4 report overall and pedestrian-specific performance across clear, snowy, and foggy conditions, using identical GCSA candidates and fixed post-processing configurations.

The results show larger observed gains in snow and fog, with particularly pronounced improvements for pedestrians. Pedestrian gains reach +0.4026 AP at 50–80 m in snow and +0.4844 AP at 30–50 m in fog. The 95% bootstrap intervals remain above zero for both overall and mean pedestrian gains across all three weather groups. On the full test set, pedestrians account for approximately 95.8% of the aggregate AP improvement. This concentration of benefits in the pedestrian category explains why averaging across all three categories yields a smaller overall increment.

Table 3. Overall 3D AP₄₀ under different weather conditions. Δ denotes the improvement over matched Gaussian Soft-NMS.

WeatherGaussianSG-GSRΔ AP95% bootstrap CI
Clear65.920165.9574+0.0372[0.0115,0.0664]
Snow65.419765.5044+0.0848[0.0595,0.1225]
Fog55.112655.2219+0.1092[0.0247,0.1743]

Table 4. Pedestrian 3D AP₄₀ averaged over the three distance ranges. Both tables use 2,000 paired sequence-level bootstrap resamples of the fixed predictions; CI denotes confidence interval.

WeatherGaussianSG-GSRΔ AP95% bootstrap CI
Clear48.075448.1837+0.1083[0.0321,0.1944]
Snow67.060467.3025+0.2421[0.1684,0.3565]
Fog63.321563.6329+0.3114[0.0633,0.5114]
AuthorsOpenReview ↗

Response to Reviewer uKhd (3/4)

2. Method contribution and SG-GSR (continued)

Reviewer comment (W2; continued).

A comparison with existing adaptive or learned suppression methods would help position SG-GSR: Adaptive NMS (Liu et al., CVPR 2019), FeatureNMS (Salscheider, 2020), learned NMS / joint rescoring (Hosang et al., CVPR 2017).

Adaptive and learned suppression. We implemented the requested Adaptive NMS [2], FeatureNMS [3], and learned NMS [4] comparisons on the same frozen GCSA candidate set. The added modules are trained for 20 epochs.

Table 5. Parameters and epochs concern only the additional post-processing module. The detector is frozen. The four trained rows all use epoch 20.

Post-processingAdditional learnable parametersTrain epochsOverall AP
Gaussian Soft-NMS0069.6588
SG-GSR0069.7193
Adaptive, BEV19,2352069.5450
Adaptive, image-plane18,3712069.5529
FeatureNMS, BEV7,821,3442069.7023
Learned NMS, BEV542,3692069.7965

Training-free post-processing. SG-GSR achieves 69.7193 Overall AP without additional post-processing training or learned parameters. On the same frozen GCSA candidates, its accuracy is comparable to the 3D adaptations of FeatureNMS (69.7023 AP) and learned NMS (69.7965 AP), both trained for 20 epochs. These results demonstrate that competitive prediction-set refinement can be achieved without training an auxiliary suppression network.

SG-GSR directly incorporates candidate-aligned gated-slice and geometric evidence into score suppression at inference. It operates on detector outputs while preserving box geometry and class labels, allowing integration with a trained detector without retraining it. This design avoids the additional optimization stage and learned parameters required by trainable post-processors, while retaining the ability to adapt suppression strength to the sensor evidence associated with competing predictions.

Comments on the paper

1 comment
AuthorsOpenReview ↗

summary of revisions and responses

Dear Area Chair and Reviewers,

Thank you for your constructive feedback. We summarize the additional evidence addressing SG-GSR's contribution, GCSA's gains under controlled computation, and the framework's performance under adverse sensing conditions. All AP differences are in 3D AP40 points.

  1. SG-GSR: improving adverse-weather pedestrian detection without additional training.

SG-GSR uses candidate-specific gated and geometric evidence to refine confidence scores, allowing suppression decisions to reflect the sensor support of competing predictions. It operates on the outputs of a trained detector while preserving box geometry and class labels.

On identical frozen GCSA candidates, SG-GSR improves mean pedestrian AP over matched Gaussian Soft-NMS by 0.2421 in snow and 0.3114 in fog, with both 95% confidence intervals above zero. Gains reach 0.4026 at 50-80 m in snow and 0.4844 at 30-50 m in fog. Across the full test set, pedestrians account for approximately 95.8% of the sensor-conditioned Overall increment (main Table 8; Supplementary Sec. I.2).

The evidence-alignment control supports the role of candidate-specific sensor information: correctly aligned evidence achieves 69.7193 Overall AP, whereas shuffling evidence among same-class candidates within each frame reduces it to 69.6677 on average, near the Gaussian baseline of 69.6588 (Supplementary Sec. B.5).

SG-GSR also achieves accuracy close to the tested 3D adaptations of FeatureNMS and learned NMS: 69.7193 versus 69.7023 and 69.7965 AP, respectively. Those auxiliary modules were trained for 20 epochs; SG-GSR requires neither additional training nor learned parameters (Supplementary Sec. H.2).

To make the gain attribution explicit, the 0.4944 Overall AP improvement over GCSA comprises 0.4339 from ordinary Gaussian decay and 0.0605 from sensor conditioning (main Table 6). For the latter, 2,000 paired bootstrap resamples over 484 acquisition sequences yield a 95% confidence interval of [0.0365, 0.0793], supporting a positive increment under sequence resampling of the fixed predictions (Supplementary Sec. I.2).

  1. GCSA gains persist under controlled comparisons.

GCSA exceeds the higher-resolution pseudo-RGB control by 1.6765 AP with approximately 0.40% difference in gated-branch MACs, and exceeds uniform aggregation within the same shared three-pass architecture by 2.0826 AP (main Tables 4-5). These comparisons support the benefit of geometry-conditioned aggregation beyond increasing encoder computation alone. With identical Gaussian post-processing, GCSA also retains a 3.4689 AP advantage over SAMFusion (Supplementary Sec. H.1). Full-detector parameter counts, computation, and latency are reported in Supplementary Sec. G.

  1. Adverse-weather performance and geometric reliability.

At the complete-system level, the model improves Overall AP over the LiDAR-RGB baseline by 2.2298 in snow and 3.0861 in fog, with mean pedestrian gains of 5.2809 and 6.7003 AP, respectively (main Table 7; Supplementary Sec. I.1).

The geometry study separately tests corrupted routing conditions, projection-path errors, and sparse projected depth. Routing-only perturbations cause a maximum observed Overall AP drop of 0.1147. GCSA retains higher absolute AP than both baselines under all tested projection-path and sparsity settings; large RGB projection errors remain a limitation (Supplementary Sec. J).

We also revised Figures 1-2 and Sections 3.1-3.3 to clarify the motivation, processing flow, and roles of feature aggregation and score refinement. The training strategy is detailed in Section 4.1 and Supplementary Sections A.2-A.4.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.