acceptodds
Under review as a conference paper at ICLR 2027

Consolidate Then Refine: 3D Object Detection with LiDAR and 4D Radar in Adverse Weather

Abstract

Adverse weather degrades LiDAR and camera observations, whereas 4D radar remains relatively stable but suffers from limited spatial resolution. Existing LiDAR–radar detectors mainly fuse the two modalities at the scene level, which couples their representations and prevents modality-specific evidence from being fully exploited. We develop a consolidate-then-refine framework that preserves modality-specific representations and associates their evidence at the object level. This framework consolidates independently generated LiDAR and radar candidates into unified object hypotheses and progressively refines their representations and geometry. On the official K-Radar v2.1 benchmark, our detector achieves 66.0% 3D [email protected], surpassing the state of the art by 9.5 points.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
3D object detection, LiDAR, 4D Radar, multimodal sensor fusion, adverse weather, autonomous driving

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_zW9ROpenReview ↗
Rating
1
Confidence
4
Informativeness
1
Presentation
1
Soundness
2
Read review
Paper summary

This paper proposes a "consolidate-then-refine" framework for 3D target detection using LiDAR-4D Radar in adverse weather conditions, which includes independent modality encoding, center candidate voting, and two-stage geometric condition decoding. The method achieved high detection results on K-Radar v2.1. However, the current paper has serious problems in writing, innovative argumentation, and experimental design. The main text fails to clearly convey the core problem solved by the paper; the main mechanism is not established in comparison with existing 3D detection methods; and the experiments are insufficient to support the main conclusion of the paper regarding robustness in adverse weather conditions.

Overall, the quality of this paper is really poor, I therefore suggest rejection.

Critical strengths
  • The research topic has practical significance, and the complementarity between LiDAR and 4D Radar is also quite clear.
  • The numerical results on K-Radar v2.1 are relatively competitive.
  • The paper provides some supplementary experiments, failure cases, and mechanism visualizations.
Critical weaknesses
  1. The readability of the paper is seriously lacking. The abstract is too short, containing only a brief background, the name of the framework, and a general result, without clearly explaining the core challenges, why the methods can solve these challenges, the main experimental findings, and the scope of application of the methods. The introduction is more like a list of related works, lacking a clear "problem - related work - challenge - motivation - method" logic. What's worse, the length of the main text and the information it carries are significantly insufficient. Much key content is compressed into supplementary materials, resulting in insufficient space for the main text. Moreover, the main text only has one framework diagram and one qualitative result diagram, lacking explanatory diagrams for central voting, the Radar encoder, two-stage sampling, and failure modes. The existing Figure 1 is generated directly by the Agent, which is too abstract. Figure 2 is a result puzzle and cannot fulfill the function of method explanation.

  2. The innovative argument is not sufficiently supported. The generation of candidate center heatmaps, initialization of candidate objects, object queries with deformable attention, and secondary feature sampling based on initial boxes all belong to relatively mature designs in 3D detection. CenterFormer has already used center candidate initialization for Transformer queries and combined it with multi-scale deformable attention. Therefore, this paper needs to clearly explain the substantial technical differences between center voting and box-conditioned refinement compared to existing methods. The current paper more presents a combination of several existing mechanisms, without demonstrating that this combination has a unique necessity for adverse weather conditions.

  3. The experimental quality is insufficient to support the main conclusion. The experiment only used one dataset, K-Radar. Although the paper acknowledged this, it is an issue that cannot be bypassed in the experimental part. The comparison method is obviously incomplete. The main text only compared PointPillars, BEVFusion, L4DR, and ASF, without including many of the latest published works. Without these methods, the paper cannot fully claim that the current results represent the state of the art.

Questions for authors

See all of Critical Weaknesses

3 comments

We thank the reviewer for the constructive comments. Alongside the improved drawing of Figure 1, we provide a direct technical comparison with CenterFormer, present evidence of radar's contribution under different weather and LiDAR visibility conditions, and clarify the evaluation protocol and comparison scope.

W1. Readability

Abstract. The abstract identifies the sensing challenge, the fusion strategy, and the main findings.

  1. LiDAR degrades in adverse weather, while radar remains more stable but has limited spatial resolution. Scene-level fusion couples their representations, limiting the use of their distinct evidence.
  2. The framework consolidates independently generated candidates into shared object hypotheses and refines their representations and geometry while preserving separate modality features.
  3. The reported result is 66.0% 3D mAP at IoU 0.5 on K-Radar v2.1, with a 9.5-point gain over the strongest baseline evaluated under our common protocol.

Introduction and main text. The introduction follows a problem–related work–challenge–motivation–method progression. Paragraphs 1–2 establish weather-induced LiDAR degradation and radar's complementary but coarse observations. Paragraphs 3–4 review local, BEV, and query-based fusion, identifying displaced evidence, coupled representations, and duplicate candidates as challenges. These motivate independent modality features and unified object hypotheses, leading to the consolidate-then-refine design in paragraph 5. Sections 3.2–3.4 detail the radar encoder, voting, and two-stage sampling.

Explanatory diagrams. The redrawn Figure 1 preserves the content and structure of our hand-drawn original. Candidate paths meet at center voting to form shared hypotheses. D1 samples around each center and predicts a preliminary box, whose geometry guides refinement in D2. Arrows from the two feature streams show that both stages read each modality's features, while the insets illustrate their sampling references. Figures C2–C3 show voting and sampling on actual scenes, and Appendix D.2 illustrates failure modes.

Figure 2 provides a qualitative comparison of detection results. Within each weather group, the rows compare methods on the same frame, and the image, radar BEV, and LiDAR BEV views show corresponding targets. The caption identifies the methods, weather groups, and prediction/annotation markers.

W2. Technical differences from prior methods

Center voting and box-conditioned refinement. CenterFormer's released single-frame model also uses two decoder layers. The differences concern candidate consolidation, feature sources, and sampling geometry.

AspectCenterFormer, released single-frame modelOur method
Query initializationCenters selected from a LiDAR heatmapIndependently generated LiDAR and radar candidates consolidated through spatial voting
Feature evidenceLiDAR featuresSeparate LiDAR and radar features read at both stages
Sampling referenceCandidate centers across decoder layers, with learned offsetsConsolidated centers in D1; the preliminary box's center and four BEV corners in D2, with learned offsets

Adverse weather conditions. When LiDAR evidence is weak, radar candidates can still support object hypotheses. Sedan AP is similar in Normal at 64.83 for LiDAR-only and 64.53 for fusion, while Heavy snow shows a gain from 53.18 to 60.78 (+7.60 points; Table 5). Among 334 Heavy-snow Sedan objects with no LiDAR points inside their ground-truth boxes, LiDAR-only, radar-only, and the full model detect 0, 97, and 104, respectively (confidence > 0.3, 3D IoU > 0.5). These results support radar's complementary role under degraded LiDAR observations (Appendices B.2–B.3).

W3. Dataset coverage and recent comparisons

K-Radar provides synchronized LiDAR, the radar power inputs required by our encoder, and seven weather categories. Evaluation on another dataset requires adapting the radar-input pipeline and establishing new training and evaluation settings. This additional work could not be completed within the revision period.

Further comparisons were limited by access to implementations and weights. Pretrained 3D-LRF weights were unavailable to us, and our attempt with the released code did not reproduce its reported results. Official REL code and weights could not be obtained, and DLR-Fusion's linked repository was unavailable during our checks. We appreciate the reviewer's understanding of these constraints.

Published AP cannot be directly reused under our protocol. The original K-Radar heading export rotates boxes by 90° relative to their physical orientation, changing IoUs and detection matches. Correcting this requires re-evaluating predictions (Appendix A.3). Under the common protocol, L4DR and ASF (NeurIPS 2025) checkpoints yield 56.47 and 52.61 mAP, versus our 65.96. The 9.49-point gain refers to this evaluated set.

Thank you for your detailed response. However, I believe the quality of the initial draft is still far from meeting the standards required for acceptance at a top-tier machine learning conference, and this gap cannot be bridged through rebuttal alone.

The authors should take time to study what constitutes a good paper and thoroughly refine their work before submission, rather than submitting an AI-generated report that suffers from poor readability.

Therefore, I maintain my original score.

Thank you for your follow-up. We acknowledge that our manuscript has room for improvement in presentation, but we disagree with your overall assessment of its quality. We do not believe the paper deserves such a sweeping dismissal. We have carefully addressed the concerns raised in your original review, yet your latest comment does not engage with our response. Combined with the generic and formulaic nature of some criticisms in the original review, this leaves us questioning whether our work has received a careful and independent reading.

Reviewer_Vqt6OpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

The paper proposes a consolidate-then-refine framework for LiDAR–4D radar 3D object detection under adverse weather. It independently generates modality-specific object-center candidates, consolidates them via cross-modal center voting, and progressively refines 3D bounding boxes using center- and box-conditioned deformable attention. On K-Radar v2.1, it achieves 65.96% 3D mAP at IoU 0.5, outperforming the strongest reported baseline by 9.49 percentage points. Ablation studies support the contributions of candidate consolidation, geometric refinement, and the radar encoder, though validation is limited to a single dataset.

Critical strengths
  1. Effective fusion strategy: Center voting consolidates misaligned LiDAR and radar candidates while preserving modality-specific features, addressing a key limitation of conventional fusion.
  2. Strong empirical validation: Achieves 65.96% 3D mAP on K-Radar v2.1, with ablations supporting the contributions of the proposed components.
Critical weaknesses
  1. Limited novelty: The method mainly combines existing techniques, with insufficiently demonstrated advances over prior object-level fusion methods.
  2. Limited validation: Evaluation on a single dataset without multi-seed experiments limits confidence in generalization and stability.
  3. Unclear attribution of gains: Even the weakest ablation (61.58%) outperforms the strongest baseline by about 5 points, suggesting gains may largely come from the backbone and training strategy.
  4. Insufficient ablations: Modality-only and no-voting baselines are missing. Progressive refinement adds only 0.58 points, and the comparable gains in normal weather weaken the adverse-weather robustness claim.
Questions for authors
  1. Could the authors provide LiDAR-only, radar-only, and no-voting baselines to clarify each component's contribution?
  2. Can the gains from the backbone, training strategy, and proposed modules be disentangled?
  3. How do the authors justify the adverse-weather robustness claim given similarly large gains in normal weather?

Suggestions for Future Work

  1. Evaluate generalization across datasets and weather conditions, including robustness curves for adverse-weather subsets.
  2. Analyze center voting sensitivity to candidate quality and radar sparsity.
2 comments

We thank the reviewer for the constructive comments. We added LiDAR-only, radar-only, independent candidate generation (no voting), and single-stage decoder baselines, an anchor baseline with the same encoder architectures, and analyses of LiDAR visibility.

Q1. LiDAR-only, radar-only, and no-voting baselines

Table 3 reports the requested comparisons using 3D AP at IoU 0.5 on K-Radar v2.1.

ModelSedanBus/TruckmAP
LiDAR-only67.3361.6264.48
Radar-only34.3826.5530.46
Independent candidate generation (no voting)66.9961.5864.29
Full model68.5063.4365.96

The single-modality models retain the corresponding encoder and both decoder stages, using candidates and features from that modality alone. The full model improves over LiDAR-only by 1.49 mAP points. The weather and visibility analyses in Q3 examine where radar contributes useful evidence.

The no-voting variant selects 32 candidates from each modality and concatenates them into 64 queries without voting. The full model starts with 64 candidates per modality and uses center voting to consolidate candidates supporting the same object into shared hypotheses, gaining 1.68 points. Both use the same two-stage decoder with a fixed capacity of 64 query slots (Appendix A.8).

Progressive refinement. The following decoder variants retain the full model's candidate generation and voting.

DecoderStages3D mAP
Single-stage decoder163.25
Center-conditioned D2 sampling265.38
Full model265.96

In the full model, D1 samples around the consolidated center and predicts a preliminary box. D2 uses this box's center and BEV corners to guide further feature sampling and refinement. The single-stage variant removes D2 and predicts final boxes directly from D1; adding the refinement stage improves mAP by 2.72 points.

The center-conditioned D2 variant keeps both stages but replaces the box-derived sampling anchors with D1's predicted center. The 0.58-point gain therefore concerns using box extent and orientation to guide sampling within the second stage (Table 3; Appendix A.8).

Q2. Backbone, training strategy, and proposed modules

Backbone and proposed modules. We compare the full model with an anchor baseline using the same LiDAR and radar encoder architectures. Table 3 reports 3D mAP at IoU 0.5.

Variant3D mAPFull-model gain
Anchor head (same encoder architectures)57.74+8.22
PointPillars-style radar encoder63.01+2.95
Full model65.96—

The anchor baseline replaces our object-level backend with BEV concatenation, local convolutional fusion, and an anchor head. With encoder architectures held fixed, the 8.22-point gap reflects the backend and its associated supervision. The no-voting and decoder controls in Q1 further examine candidate consolidation and refinement.

The radar-encoder control replaces the ordered vertical power representation with PointPillars-style aggregation while retaining voting and both decoder stages. Its 2.95-point gap measures the effect of this encoder change.

Training strategy. The anchor baseline and full model are trained from scratch with the same split, augmentation, and optimization schedule. Both use 20 epochs, AdamW, an initial learning rate of 2×10⁻⁴, and batch size four (Appendix A.8). These comparisons hold the training recipe fixed; we have not separately quantified gains from changing the schedule or augmentation.

Q3. Adverse-weather robustness

Normal and adverse weather. Margins over external detectors reflect differences in encoders, fusion, and detection heads. To examine radar's contribution under weather degradation, we compare the full model with its LiDAR-only variant (Table 5; Table B3). Sedan AP changes from 64.83 to 64.53 in Normal but rises from 53.18 to 60.78 in Heavy snow, a gain of 7.60 points. Bus/Truck gains are 2.25 points in Normal and 2.11 in Heavy snow.

The benefit varies across weather and classes. In Light snow, AP drops by 0.19 points for Sedan and 5.15 for Bus/Truck; Bus/Truck AP also drops by 0.79 points in Sleet. Our evidence for weather robustness is the ability to use radar when LiDAR observations deteriorate, particularly for Sedan in Heavy snow.

Sensor visibility. The following table counts detected Heavy-snow objects with no LiDAR points inside their ground-truth boxes (confidence > 0.3, 3D IoU > 0.5):

ClassGround-truth boxesLiDAR-onlyRadar-onlyFull model
Sedan334097104
Bus/Truck2432136140

Radar provides detection evidence for these objects despite the absence of direct LiDAR returns within their boxes. Appendix B.3 extends the analysis to different LiDAR point-count groups and reports both recall and false-positive counts.

W1. Prior object-level fusion methods

TransFusion predicts boxes from LiDAR queries before reading image features. SparseFusion generates candidates in separate branches and concatenates them in a common 3D space for self-attention fusion. Our voting field instead associates spatially displaced candidates before decoding, producing hypotheses supported by either or both sensors. Each hypothesis initializes a shared query center and representation, while the LiDAR and radar features remain separate for both decoder stages to read (Sections 3.3–3.4).

The contribution lies in how cross-modal candidates are associated and how the resulting geometry guides feature retrieval. This is relevant when the two sensors provide partial, spatially displaced responses to the same object. The no-voting and refinement controls in Q1 examine these choices within the same detection framework.

W2. Generalization and stability

Evaluation on a single dataset. We chose K-Radar because it provides synchronized LiDAR and the 4D radar power observations used by our encoder, together with seven weather categories. Evaluation on another dataset requires a compatible radar-input pipeline and a new training and evaluation setup. We appreciate this suggestion, but have not been able to complete the additional dataset study for this revision.

Multi-seed experiments. We agree that multiple training runs would provide stronger evidence for the consistency of the component gains. This requires matched repetitions of the full model and the corresponding ablations, which we have not been able to complete within the rebuttal period.

To allow verification of the reported metrics, the supplementary archive provides full-model predictions for all 13,727 test frames, evaluation code, test splits, and instructions for recomputing the results.

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

This paper proposes a consolidate-then-refine framework for LiDAR–4D radar 3D object detection under adverse weather conditions. Separate encoders extract modality-specific features, while object-centric center voting consolidates LiDAR and radar candidates into unified object hypotheses. A two-stage geometry-conditioned decoder progressively refines object representations and bounding boxes. The method also introduces a radar encoder designed to preserve vertical power distributions. Experiments on K-Radar v2.1 demonstrate improved detection performance across multiple weather conditions. The main contribution is an object-level fusion framework that preserves modality-specific representations and progressively aggregates complementary sensor information.

Critical strengths

1.The object-level fusion strategy is well motivated. By independently extracting LiDAR and radar features and consolidating their object candidates, the method addresses spatial misalignment between different sensor observations. The progressive refinement mechanism provides a reasonable approach to improving 3D localization accuracy. 2.The experimental results demonstrate substantial performance improvements on K-Radar v2.1, particularly under the stricter 3D IoU threshold. The paper includes component ablations, weather-specific evaluations, and qualitative visualizations. These results provide useful evidence for the effectiveness of the proposed framework.

Critical weaknesses

1.The physical complementarity between LiDAR and 4D radar is not sufficiently demonstrated. Although the paper emphasizes the advantages of radar under adverse weather, it does not clearly establish which missing or degraded LiDAR information is compensated for by radar. The existing visualizations illustrate cross-modal candidate association but provide limited quantitative evidence of actual information compensation. More targeted modality-specific evaluations, particularly under different sensor visibility conditions, would help clarify the necessity and effectiveness of the proposed multimodal fusion. 2.The ablation studies do not fully isolate the contributions of the proposed components. In Table 3, the fused-heatmap variant simultaneously replaces modality-specific candidate generation and center voting, making their individual effects difficult to distinguish. Moreover, the improvement from box-conditioned refinement over center-conditioned sampling is relatively small, and a single-stage decoder baseline is missing. Additional controlled experiments are needed to establish the independent contributions of center voting and progressive refinement. 3.The comparisons do not adequately cover recent LiDAR–radar detection advances. Relevant methods such as DLR-Fusion (ICCV 2025) and REL (AAAI 2026), both discussed in the manuscript, are not included in the experimental comparisons. Evaluations against these methods under consistent input settings and protocols would provide stronger evidence for the claimed superiority. 4.The robustness and practical efficiency of the proposed framework require further evaluation. Although the method performs well across weather conditions, its Bus/Truck detection performance remains limited in some challenging subsets. In addition, the paper does not report inference latency, computational complexity, or sensitivity to sensor calibration errors. These evaluations would help establish the reliability and practicality of the framework for autonomous driving applications.

Questions for authors

1.What specific information does 4D radar provide to compensate for degraded LiDAR observations? Can the authors provide LiDAR-only, radar-only, and fused-model comparisons under different weather and sensor visibility conditions to demonstrate the actual complementary benefits? 2.Can the authors separately evaluate modality-specific candidate generation, center voting, and progressive refinement? In particular, how does the proposed decoder compare with a single-stage alternative, and how consistent are the improvements across multiple training runs? 3.Can the authors include comparisons with more recent LiDAR–radar fusion methods, such as DLR-Fusion and REL, under consistent evaluation settings? Can inference latency and robustness to sensor misalignment also be reported?

2 comments

We thank the reviewer for the constructive comments. We added LiDAR-only, radar-only, independent candidate generation (no voting), and single-stage decoder baselines, together with analyses of LiDAR visibility and inference efficiency.

Q1. Radar compensation for degraded LiDAR observations

Radar provides object-location cues through spatially distributed power returns when an object has few or no direct LiDAR returns.

Modality comparisons. The single-modality variants retain the corresponding encoder and the two-stage decoder, with both stages reading only that modality's features (Appendix A.8). Table 3 reports 3D mAP at IoU 0.5 on K-Radar v2.1:

ModelmAP
LiDAR-only64.48
Radar-only30.46
Full model65.96

Fusion improves on LiDAR-only by 1.49 mAP points overall. In Heavy snow, Sedan AP rises from 53.18 to 60.78 (+7.60 points) and Bus/Truck AP from 36.84 to 38.95 (+2.11 points) (Table 5; Table B3). Radar-only AP is 42.65 and 20.42, respectively.

Sensor visibility. We also group objects by the number of LiDAR points inside their ground-truth boxes. For the zero-point group across the full test set, the three models have the following recall (confidence > 0.3, 3D IoU > 0.5):

ClassBoxesLiDAR-onlyRadar-onlyFull model
Sedan1,0691.59%18.15%22.36%
Bus/Truck2760.72%49.28%50.72%

These results show that radar contributes detection evidence for objects lacking direct LiDAR returns. Appendix B.3 reports recall and false-positive counts across LiDAR point-count groups.

Q2. Candidate generation, center voting, and refinement

Candidate generation and center voting. We added an independent candidate generation baseline that retains the separate modality heatmaps but omits voting. Table 3 now compares:

Variant3D mAP
Fused-heatmap initialization61.58
Independent candidate generation (no voting)64.29
Full model65.96

The fused-heatmap variant predicts a shared heatmap from concatenated LiDAR and radar BEV features. Independent candidate generation instead selects 32 candidates from each modality's heatmap and concatenates them into 64 queries. Keeping separate candidate generation raises mAP by 2.71 points over the fused-heatmap variant.

The full model starts with 64 candidates per modality. Center voting consolidates candidates supporting the same object into a shared hypothesis, which initializes a query's center and representation. It gains 1.68 points over independent candidate generation. All three variants use the same two-stage decoder with a fixed capacity of 64 query slots (Appendix A.8).

Single-stage alternative. These variants retain the full model's candidate generation and center voting, and change the decoder (Table 3; Appendix A.8):

DecoderStages3D mAP
Single-stage decoder163.25
Center-conditioned D2 sampling265.38
Full model265.96

In the full model, D1 samples around the consolidated center and predicts a preliminary box. D2 uses this box's center and BEV corners to guide further feature sampling and box refinement. The single-stage variant removes D2 and predicts final boxes directly from D1; adding the refinement stage improves mAP by 2.72 points.

The center-conditioned D2 variant keeps both stages but replaces the box-derived sampling anchors with D1's predicted center. The full model gains 0.58 points by using the box's extent and orientation to guide sampling.

Multiple training runs. We agree that multiple training runs would provide stronger evidence for the consistency of the component gains. This requires matched repetitions of the full model and the corresponding ablations, which we have not been able to complete within the rebuttal period. We would appreciate the reviewer's understanding.

To allow verification of the reported metrics, the supplementary archive provides full-model predictions for all 13,727 test frames, evaluation code, test splits, and instructions for recomputing the results.

Q3. Recent methods, latency, and sensor misalignment

DLR-Fusion and REL. We agree that these recent methods are relevant comparisons. REL has not released code or weights, and DLR-Fusion's linked repository is unavailable (as of November 13, 2026). These limitations have prevented us from completing their evaluation under our protocol.

ASF (NeurIPS 2025) is a recent strong K-Radar baseline. On benchmark v1.0 (lateral ROI ±6.4 m), its camera–LiDAR–radar model reports 73.6 Sedan 3D AP at IoU 0.5, compared with 45.2 for 3D-LRF, 53.5 for L4DR, 54.5 for REL, and 45.7 for DLR-Fusion [ASF, Table 1; REL, Table 1; DLR-Fusion, Table 1].

Published AP values cannot be directly reused under our protocol: the original heading export rotates boxes by 90° relative to their physical orientation, changing IoUs and detection matches. We correct this conversion and re-evaluate predictions (Appendix A.3). Under our common K-Radar v2.1 protocol, released L4DR and ASF checkpoints yield 56.47 and 52.61 two-class 3D mAP, versus our 65.96, gains of 9.49 and 13.35 points. ASF also uses camera input.

3D-LRF, REL, and DLR-Fusion focus on interaction before BEV formation; L4DR and ASF also fuse BEV features. Our method independently encodes the two modalities, then consolidates object candidates and refines their boxes. The results indicate that fusion at this later stage can make effective use of separately extracted features.

Object-level fusion has been studied more extensively in LiDAR–camera detection, including TransFusion and SparseFusion. Adapting these methods to LiDAR–radar requires changing modality encoders and cross-modal association, so their original performance advantages may not carry over. We include a LiDAR–radar BEVFusion adaptation and controlled component comparisons to examine our design choices on K-Radar (Table 3; Appendix A.8).

Inference latency. Our model takes 30.88 ms/frame (32.39 FPS), with 11.74M parameters and peak GPU memory of 0.40 GiB allocated and 0.66 GiB reserved. Appendix B.4 compares all main baselines using their main-result checkpoints and configurations on an RTX 4090 and Intel Core Ultra 9 285K, with FP32, batch size one, and two CPU threads. Each method processes the same 2,000 test frames: five for warm-up and 1,995 for averaging latency.

Timing starts with sensor inputs in CPU memory and includes voxelization, input feature construction and collation, GPU transfer, network inference, and detection postprocessing. GPU execution is synchronized; disk I/O, calibration, cropping, and image preprocessing precede timing. FPS is 1000 divided by mean latency in milliseconds. LiDAR voxelization takes 1.40 ms and radar column feature construction takes 6.66 ms per frame, together accounting for 26.1% of our model's latency.

Sensor misalignment. Our experiments evaluate weather degradation under the dataset's provided calibration. Quantifying sensitivity to misalignment requires an additional evaluation with controlled translation and rotation perturbations, which we have not been able to complete for this revision. Appendix E discusses how calibration drift can separate corresponding center candidates and displace the feature samples used for refinement.

W4. Bus/Truck performance

Sparse training support can limit Bus/Truck detection in some subsets. Rain and Overcast contain only 179 and 106 training boxes, respectively (Table A2). In the Overcast test subset, 132 of 151 boxes come from one sequence, making AP sensitive to individual errors and scene composition. In Heavy snow, 91.17 BEV AP versus 38.95 3D AP is consistent with vertical localization errors (Appendix B.2), illustrated by the 1.29 m vertical-center error in Appendix D.2.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.