SlimComm: Doppler-Guided Sparse Queries for Bandwidth-Efficient Cooperative 3-D Perception

Iramm Hamdard; J. Marius Zoellner; Melih Yazgan; Qiyuan Wu; Shiqi Li

REVIEW 4 major objections 4 minor 36 references

SlimComm: Doppler-Guided Sparse Queries for Bandwidth-Efficient Cooperative 3-D Perception

T0 review · 4 major / 4 minor · reviewed 2026-08-15 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗ Review JSON

Pith's one-line read SlimComm's Doppler-guided sparse query scheme shares only small feature patches, cutting cooperative-perception bandwidth by up to 90 percent while matching full-map detection accuracy.

desk verdict The accuracy-bandwidth claim is undercut by an unfair sensor comparison, but the Doppler-guided query idea and the two new datasets are genuine contributions worth peer review. read the letter →

arxiv 2508.13007 v1 pith:EUHFR4BY submitted 2025-08-18 cs.CV

Melih Yazgan , Qiyuan Wu , Iramm Hamdard , Shiqi Li , J. Marius Zoellner This is my paper · ORCID

classification cs.CV

keywords collaborativeperception4DradarDopplersparsequeriesbandwidthefficiencyBEVfeaturecommunicationocclusion-aware3Dobjectdetection

verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

SlimComm aims to show that connected vehicles can cooperate over limited V2X bandwidth without abandoning accuracy: instead of exchanging dense bird's-eye-view feature maps, each agent broadcasts a small set of queries, and only the requested feature patches are returned. The paper's central proposal is to let 4D radar Doppler decide where queries go—its ego-motion-compensated signal separates moving from static objects, while a confidence prior plants extra queries in occlusion shadows. The authors release two simulated benchmarks, OPV2V-R and Adver-City-R, with per-point Doppler radar, and report that SlimComm consumes up to 90 percent less bandwidth than full-map sharing while matching or exceeding prior cooperative perception methods in average precision. If the results hold, bandwidth-efficient cooperative perception becomes a query-selection problem rather than a compression problem, with motion and occlusion semantics doing the selection.

What carries the argument

The load-bearing mechanism is the Doppler-guided query generator. Ego-motion compensation turns radar returns into a binary motion prior: moving objects carry a nonzero Doppler residual, static background does not, so the map is motion-centric rather than confidence-only. Two query types are produced: reference queries selected on dynamic or high-confidence cells, and exploratory queries sampled in occlusion shadows cast by strong confidence peaks. A two-stage offset—a coarse nudge followed by a 3×3 deformable halo—turns each anchor into a compact sampling cluster, and the gated multi-scale deformable attention block fuses the returned patches. The mechanism's job is to concentrate the exchanged bits where evidence is strongest or likely hidden, leaving static background unsampled.

What would settle it

Equip a test fleet with per-point Doppler radar and run SlimComm on recorded scenes with dense moving and static traffic; if the motion map routes queries to static clutter or misses moving objects, the reported accuracy at low bandwidth will not reproduce.

Watch

Extended reading notes

Core claim

The central discovery is that motion and occlusion semantics, read off per-point radar Doppler, can drive a sparse communication protocol that keeps dense-map accuracy. SlimComm builds a motion-centric dynamic map from ego-motion-compensated Doppler returns, places reference queries on dynamic cells and high-confidence foreground cells, and places exploratory queries behind confidence peaks so collaborators return context for partially or fully occluded regions. Neighbours warping their features into the ego frame send only a 3×3 halo-enriched patch per query, and a multi-scale gated deformable attention block fuses the responses. On the released OPV2V-R and Adver-City-R datasets the scheme reports up to 90% bandwidth reduction relative to full-map sharing, with AP@0.5 and AP@0.7 matching or exceeding strong baselines, and degrades by less than 1.5 mAP under 200–600 ms asynchronous delays.

Load-bearing premise

The simulated per-point radar Doppler reproduces real 4D radar behavior—velocity on moving objects, near-zero Doppler on static ones—so the query placement learned in simulation will transfer to physical vehicles.

Editorial extensions

If this is right

Bandwidth per frame drops to roughly 10% of full-map sharing: on OPV2V-R, SlimComm uses 0.63 MB/frame versus 4.45 MB for the next most efficient baseline and tens of MB for dense-map methods.
Accuracy is not traded away: on Adver-City-R SlimComm reaches AP@0.5 of 0.67 and AP@0.7 of 0.54 in the general split, at or above AttFusion and S-AdaFusion, and on OPV2V-R it stays competitive while using 7x less bandwidth than Where2Comm.
Exploratory queries materially help occluded cases: ablations show the exploratory branch lifts AP@0.7 on Adver-City-R from 0.41 to 0.49, and the halo enrichment lifts it further to 0.54.
Stability under pose and timing errors: SlimComm maintains the best AP@0.7 across localization noise up to 0.6 m and heading noise up to 1.0°, and loses less than 1.5 mAP under delays up to 600 ms.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

Editorial inference: the same query logic should transfer to setups without radar by substituting optical flow or learned motion cues, since the essential signal is a cheap motion-versus-static prior rather than radar specifically.
Editorial inference: on real roads with many stationary objects, the bandwidth saving should grow with traffic density, because queries concentrate on movers and occlusion shadows rather than background; a real-radar benchmark could test this.
Editorial inference: the query budget itself could be tuned dynamically from scene statistics, for example raising the exploratory-query count at intersections where occluders are dense.

Share X Bluesky LinkedIn Reddit HN

Signed reviews

No signed human review yet.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Desk Editor's Note

read the letter

You should know two things about this paper. First, the core idea is genuinely new: using Doppler radar to place sparse queries in cooperative perception, plus an occlusion-shadow branch that probes behind occluders. That's a real step beyond the confidence/entropy heuristics in Where2Comm and EffiComm. Second, the headline comparison in Table 2 is apples-to-oranges. SlimComm is trained with LiDAR plus Doppler radar; the baselines (AttFusion, S-AdaFusion, SCOPE, Where2Comm) are LiDAR-only methods with no radar channel. No passage says the baselines were adapted to consume radar, and there is no 'SlimComm without radar' ablation. So the matching-or-better accuracy at ~10% bandwidth may reflect the extra sensor modality, not the sparse-query scheme. That is the load-bearing gap.

The paper also does some things well. The ablations are honest: they show the exploratory branch and halo enrichment each add a few AP points, and they report bandwidth in both log scale and MB/frame. The datasets, OPV2V-R and Adver-City-R, are a real contribution—CARLA-based with per-point Doppler, and the paper is upfront that they are simulated. The robustness experiments under localization noise and async delays are useful, though the delay curves are relegated to the supplement.

The soft spots are three, in decreasing importance. First, the sensor-fairness problem above. If the method is meant to be a full LiDAR+radar system, then compare it against LiDAR+radar baselines, or at least ablate the radar. Second, simulated radar realism: C-Shenron is a plausible simulator, but no evidence is given that its Doppler behavior transfers to real 4D radar, so the motion-prior benefits could be simulation artifacts. Third, the exact query budget and offset parameters are missing, and the code/data are promised but not released. That's minor if they do release, but it makes reproduction currently impossible.

Overall: this is a solid systems paper with an interesting idea and useful datasets, but the central empirical claim is not yet supported. It deserves a serious referee—the novelty is real and the datasets have standalone value—but the revision should require a radar-less SlimComm variant and at least one radar-augmented baseline. I'd send it to review, not desk reject, and tell the authors exactly what's missing.

Referee Report

4 major / 4 minor

Summary. SlimComm is a cooperative 3-D perception framework that uses 4D radar Doppler to construct a motion-centric dynamic map and generate two types of sparse queries: reference queries on dynamic or high-confidence regions and exploratory queries into occluded areas. Only query-specific BEV features are exchanged and fused through a gated multi-scale deformable attention module. The authors introduce OPV2V-R and Adver-City-R, CARLA-based extensions of existing V2X suites with per-point Doppler radar, and report up to 90% lower bandwidth than full-map sharing while maintaining competitive AP@0.5 and AP@0.7 against AttFusion, S-AdaFusion, SCOPE, and Where2Comm. The paper also evaluates robustness to localization noise, heading noise, and asynchronous delays up to 600 ms.

Significance. If validated, the Doppler-guided sparse-query idea is a useful contribution to bandwidth-efficient V2X perception, and the release of radar-augmented CARLA datasets could support further research on multi-agent 4D radar fusion. The paper provides clear ablations for the exploratory-query and halo-enrichment components, and the code and dataset release are valuable for reproducibility. However, the central empirical claim is not yet fully supported because the main comparison is confounded by sensor modality and the evaluation relies entirely on synthetic radar with no real-world validation.

major comments (4)

[Section 5.2.1, Table 2] The headline comparison is not apples-to-apples: SlimComm fuses LiDAR with 4D radar (with Doppler), while AttFusion, S-AdaFusion, SCOPE, and Where2Comm are LiDAR-only cooperative-perception methods. The manuscript does not state that these baselines were modified to consume the radar points or the Doppler channel. Consequently, the reported accuracy parity/superiority and the apparent robustness of query placement could stem from the additional sensor modality rather than from the sparse-query communication scheme. To support the central claim, please add a LiDAR-only variant of SlimComm as a control and, ideally, radar-augmented versions of the baselines so that all methods receive the same sensor inputs.
[Section 5.3, Table 3] The ablation study in Table 3 decomposes the full model by removing exploratory queries and halo enrichment, but it never removes the radar input or the Doppler-derived dynamic map. Without a "SlimComm without radar" condition, the improvement over the LiDAR-only baselines cannot be attributed to the Doppler-guided query mechanism. Please add such an ablation, and if feasible an ablation with radar but without Doppler gating, to isolate the contribution of the proposed sparse-query scheme from the contribution of the extra sensor modality.
[Section 3 and Table 1] All experiments are conducted on CARLA-based datasets with radar data generated by the C-Shenron simulator, and the manuscript does not provide any validation of the simulated per-point Doppler against real 4D radar measurements. Since the method's query placement and motion-centric filtering rely on the Doppler channel, the transferability of the reported results to real radar is unverified. Please state this limitation explicitly and, if possible, include a small-scale real-radar experiment or an analysis of simulator-to-real gaps.
[Section 5.2.1, Table 2] The communication cost is reported as "CV (measured in log2 scale)" and "BD (in MB/frame)", but the text does not define the exact computation: per-query payload size, halo dimensions, feature bit depth or quantization, number of collaborating agents, or averaging over frames. Without this protocol, the central quantitative claim of "up to 90% lower bandwidth" cannot be independently reproduced. Please provide the precise bandwidth formula and the query-budget, halo, and offset parameters used for the results.

minor comments (4)

[Abstract and Table 2] On OPV2V-R, SlimComm's AP@0.5 is 0.87 versus S-AdaFusion's 0.91, so the claim of "matching or surpassing prior baselines" is slightly overstated; please rephrase to "comparable or better" or report statistical significance.
[Table 2] The Adver-City-R columns mix three scenario-specific values for AP and only two scalar values for CV and BD; the presentation makes it difficult to determine which CV/BD value corresponds to which scenario. Please use separate rows or columns.
[Section 5.2.2] The text states "less than 1.5 mAP degradation" while the figure and metric are AP; please keep the metric name consistent.
[Section 3 and Supplementary Material] The C-Shenron simulator is cited, but no specific simulation parameters are given for the radar noise model or Doppler computation; please include these details so that the dataset generation is reproducible.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity; the accuracy–bandwidth claim rests on external-benchmark comparisons and visible component ablations.

full rationale

SlimComm's headline claim is an empirical accuracy–bandwidth trade-off: it reports AP@0.5/AP@0.7, communication volume, and bandwidth against four previously published collaborative-perception baselines on radar-augmented versions of OPV2V and Adver-City (Table 2). No equation in the paper defines a predicted quantity in terms of a fitted parameter, and no uniqueness or theoretical derivation is imported from the authors' prior work. The Doppler-guided query generation is a design choice tested by ablation: introducing the exploratory branch raises AP@0.5 from 0.58 to 0.61 and AP@0.7 from 0.41 to 0.49, and halo enrichment adds further gains; query-budget effects are also ablated (Tables 3–4). The only self-citations ([30], [31], [32]) appear in related-work context and are not load-bearing, while the radar sensor model [14] is an external CARLA simulation framework. The datasets are author-generated extensions of standard benchmarks, but evaluation against held-out external baselines and visible component ablations prevents the central comparison from reducing to the method's own inputs. The skeptic's concern that baselines may receive LiDAR only while SlimComm receives LiDAR plus radar is a comparison-validity/correctness issue, not a circular derivation; no quoted equation or construction exhibits the claimed result as equivalent to an input.

Assumptions & free parameters 3 free parameters · 3 assumptions · 2 invented entities

The method itself introduces no new physical entities. The main unverified inputs are the simulated radar datasets and the chosen hyperparameters for query placement and halos. The query budget and exploratory offset parameters are tuned, but their values are not reported in the provided text.

free parameters (3)

Query budget or number of reference points = Not stated in the provided text
Ablations in Section 5.3 vary the query count, suggesting the final budget is a chosen hyperparameter that trades accuracy and bandwidth.
Exploratory query shadow length and offset ranges = Not stated in the provided text
The two-stage offset and shadow sampling parameters determine how far exploratory queries reach into occluded regions.
Halo size and gating thresholds = 3 by 3 halo; thresholds not stated
The halo width is fixed at 3 by 3, and the confidence thresholds for selecting reference queries are not listed.

assumptions (3)

domain assumption CARLA and C-Shenron generate physically plausible 4D radar Doppler returns.
The whole dataset contribution and motion-centric query design assume simulated radar noise and Doppler magnitudes match real radar behavior (Section 3, Table 1).
standard math Standard PointPillars / SECOND-style LiDAR encoding is an acceptable backbone for BEV feature extraction.
The framework reuses established point-cloud encoders, which is background knowledge rather than a new assumption.
domain assumption V2X bandwidth accounting treats transmitted query features as the dominant cost and ignores negotiation or metadata overhead.
Bandwidth numbers in Table 2 count the shared feature payload, but broadcast query messages and protocol overhead are not itemized.

invented entities (2)

OPV2V-R Doppler radar dataset
purpose: Benchmark for radar-augmented cooperative perception with per-point Doppler.
The dataset is generated by the authors for this paper. Its realism is asserted, not verified against real radar measurements.
Adver-City-R Doppler radar dataset
purpose: Benchmark for radar-augmented cooperative perception under dense, occlusion-heavy traffic.
Same status as OPV2V-R; generated by the authors and not yet independently validated.

how reviews work

0 comments

Cite this review

Pith. "Pith review of SlimComm: Doppler-Guided Sparse Queries for Bandwidth-Efficient Cooperative 3-D Perception." pith.science (2026). https://pith.science/paper/EUHFR4BY

@misc{pith2026250813007,
  author       = {Pith},
  title        = {Pith review of: SlimComm: Doppler-Guided Sparse Queries for Bandwidth-Efficient Cooperative 3-D Perception},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/EUHFR4BY}},
  note         = {Machine review of arXiv:2508.13007}
}

read the original abstract

Collaborative perception allows connected autonomous vehicles (CAVs) to overcome occlusion and limited sensor range by sharing intermediate features. Yet transmitting dense Bird's-Eye-View (BEV) feature maps can overwhelm the bandwidth available for inter-vehicle communication. We present SlimComm, a communication-efficient framework that integrates 4D radar Doppler with a query-driven sparse scheme. SlimComm builds a motion-centric dynamic map to distinguish moving from static objects and generates two query types: (i) reference queries on dynamic and high-confidence regions, and (ii) exploratory queries probing occluded areas via a two-stage offset. Only query-specific BEV features are exchanged and fused through multi-scale gated deformable attention, reducing payload while preserving accuracy. For evaluation, we release OPV2V-R and Adver-City-R, CARLA-based datasets with per-point Doppler radar. SlimComm achieves up to 90% lower bandwidth than full-map sharing while matching or surpassing prior baselines across varied traffic densities and occlusions. Dataset and code will be available at: https://url.fzi.de/SlimComm.

Review history (2 revisions) →

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

36 extracted references · 31 canonical work pages

[1]

Ahmed, Siegfried Mercelis, and Ali Anwar

Ahmed N. Ahmed, Siegfried Mercelis, and Ali Anwar. Delawarecol: Delay aware collaborative perception.IEEE OpenJournalofVehicularTechnology ,6:1164–1177,2025. 2

work page 2025
[2]

IgalBilik.Comparativeanalysisofradarandlidartechnologies for automotive applications.IEEE Intelligent Transportation Systems Magazine, 15(1):244–269, 2023. 1

work page 2023
[3]

Cooper: Cooperative perception for connected autonomous vehicles based on 3d point clouds

Qi Chen, Sihai Tang, Qing Yang, and Song Fu. Cooper: Cooperative perception for connected autonomous vehicles based on 3d point clouds. In2019 IEEE 39th International ConferenceonDistributedComputingSystems(ICDCS) ,pages 514–524, 2019. 2

work page 2019
[4]

CARLA: An open urban driving simulator

Alexey Dosovitskiy, German Ros, Felipe Codevilla, Antonio Lopez, and Vladlen Koltun. CARLA: An open urban driving simulator. InProceedings of the 1st Annual Conference on Robot Learning, pages 1–16, 2017. 2

work page 2017
[5]

Deep residual learning for image recognition.arXiv preprint arXiv:1512.03385, 2015

Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition.arXiv preprint arXiv:1512.03385, 2015. 3

arXiv 2015
[6]

Where2comm: Communication-efficient collaborative perceptionviaspatialconfidencemaps

Yue Hu, Shaoheng Fang, Zixing Lei, Yiqi Zhong, and Siheng Chen. Where2comm: Communication-efficient collaborative perceptionviaspatialconfidencemaps. In AdvancesinNeural Information Processing Systems, 2022. 2, 5, 6, 7

work page 2022
[7]

V2x-r: Coop- erative lidar-4d radar fusion with denoising diffusion for 3d object detection

XunHuang,JinlongWang,QimingXia,SihengChen,Bisheng Yang, Xin Li, Cheng Wang, and Chenglu Wen. V2x-r: Coop- erative lidar-4d radar fusion with denoising diffusion for 3d object detection. InProceedings of the Computer Vision and PatternRecognitionConference(CVPR) ,pages27390–27400,

work page
[8]

Adver-City: Open-Source Multi-Modal Dataset for Collaborative Perception Under Ad- verse Weather Conditions, 2025

Mateus Karvat and Sidney Givigi. Adver-City: Open-Source Multi-Modal Dataset for Collaborative Perception Under Ad- verse Weather Conditions, 2025. arXiv:2410.06380 [cs]. 2, 6

arXiv 2025

Show all 36 references

[9]

Lang, Sourabh Vora, Holger Caesar, Lubing Zhou, Jiong Yang, and Oscar Beĳbom

Alex H. Lang, Sourabh Vora, Holger Caesar, Lubing Zhou, Jiong Yang, and Oscar Beĳbom. Pointpillars: Fast Encoders for Object Detection from Point Clouds. In2019 IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12689–12697, 2019. 3

2019
[10]

Rcbevdet: Radar-camera fusion in bird’s eye viewfor3dobjectdetection

Zhiwei Lin, Zhe Liu, Zhongyu Xia, Xinhao Wang, Yongtao Wang, Shengxiang Qi, Yang Dong, Nan Dong, Le Zhang, and Ce Zhu. Rcbevdet: Radar-camera fusion in bird’s eye viewfor3dobjectdetection. In ProceedingsoftheIEEE/CVF Conference on Computer Vision and Pattern Recognition, pages...

2024
[11]

When2com: Multi-agentperceptionviacommunication graph grouping

Yen-Cheng Liu, Junjiao Tian, Nathaniel Glaser, and Zsolt Kira. When2com: Multi-agentperceptionviacommunication graph grouping. InProceedings of the IEEE/CVF Conference oncomputervisionandpatternrecognition ,pages4106–4115,
[12]

Rus, and Song Han

Zhĳian Liu, Haotian Tang, Alexander Amini, Xinyu Yang, Huizi Mao, Daniela L. Rus, and Song Han. Bevfusion: Multi-task Multi-sensor Fusion with Unified Bird’s-eye View Representation. In2023 IEEE International Conference on Robotics and Automation (ICRA), pages 2774–2781, 2023. 1

2023
[13]

Robust collaborative 3d object detection in presence of pose errors

Yifan Lu, Quanhao Li, Baoan Liu, Mehrdad Dianati, Chen Feng, Siheng Chen, and Yanfeng Wang. Robust collaborative 3d object detection in presence of pose errors. In2023 IEEE InternationalConferenceonRoboticsandAutomation(ICRA) , pages 4812–4818. IEEE, 2023. 5

2023
[14]

DemoAbstract: C-Shenron: ARealistic 11 Radar Simulation Framework for CARLA, page 726–727

Pushkal Mishra, Satyam Srivastava, Jerry Li, Kshitiz Bansal, andDineshBharadia. DemoAbstract: C-Shenron: ARealistic 11 Radar Simulation Framework for CARLA, page 726–727. Association for Computing Machinery, New York, NY, USA,
[15]

Ro- bust multimodal vehicle detection in foggy weather using complementary lidar and radar signals

Kun Qian, Shilin Zhu, Xinyu Zhang, and Li Erran Li. Ro- bust multimodal vehicle detection in foggy weather using complementary lidar and radar signals. In2021 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 444–453, 2021. 1

2021
[16]

Adaptive feature fusion for cooperative perception using lidar point clouds

Donghao Qiao and Farhana Zulkernine. Adaptive feature fusion for cooperative perception using lidar point clouds. In Proceedings of the IEEE/CVF winter conference on appli- cations of computer vision, pages 1186–1195, 2023. 2, 5, 7

2023
[17]

PVR- CNN++: Point-Voxel Feature Set Abstraction With Local Vector Representation for 3D Object Detection.Int

Shaoshuai Shi, Li Jiang, Jiajun Deng, Zhe Wang, Chaoxu Guo, Jianping Shi, Xiaogang Wang, and Hongsheng Li. PVR- CNN++: Point-Voxel Feature Set Abstraction With Local Vector Representation for 3D Object Detection.Int. J. Com- put. Vision, 131, 2022. 5

2022
[18]

Lirafu- sion: Deep adaptive lidar-radar fusion for 3d object detection

Jingyu Song, Lingjun Zhao, and Katherine A Skinner. Lirafu- sion: Deep adaptive lidar-radar fusion for 3d object detection. In 2024 IEEE International Conference on Robotics and Automation (ICRA), 2024. 2

2024
[19]

Interfusion: Interaction-based 4D Radar and LiDAR Fusion for 3D Object Detection

Li Wang, Xinyu Zhang, Baowei Xv, Jinzhao Zhang, Rong Fu, XiaoyuWang,LeiZhu,HaibingRen,PingpingLu,JunLi,and Huaping Liu. Interfusion: Interaction-based 4D Radar and LiDAR Fusion for 3D Object Detection. In2022 IEEE/RSJ International Conference on Intelligent Robots and Systems ...

2022
[20]

V2vnet: Vehicle- to-vehicle communication for joint perception and prediction

Tsun-HsuanWang,SivabalanManivasagam,MingLiang,Bin Yang, Wenyuan Zeng, and Raquel Urtasun. V2vnet: Vehicle- to-vehicle communication for joint perception and prediction. InComputer Vision–ECCV 2020: 16th European Conference, Glasgow, UK, August 23–28, 2020, Proceedings, Part II...

2020
[21]

Bi- LRFusion: Bi-directional lidar–radar fusion for 3d dynamic object detection

YingjieWang,JiajunDeng,YaoLi,JinshuiHu,CongLiu,Yu Zhang, Jianmin Ji, Wanli Ouyang, and Yanyong Zhang. Bi- LRFusion: Bi-directional lidar–radar fusion for 3d dynamic object detection. In Proc. IEEE/CVF Conf. on Computer Vision and Pattern Recognition (CVPR), pages 13394–13404,
[22]

Rpfa-net: a 4d radar pillar feature attention network for 3d object detection

Baowei Xu, Xinyu Zhang, Li Wang, Xiaomei Hu, Zhiwei Li, Shuyue Pan, Jun Li, and Yongqiang Deng. Rpfa-net: a 4d radar pillar feature attention network for 3d object detection. In2021IEEEInternationalIntelligentTransportationSystems Conference (ITSC), pages 3061–3066, 2021. 1

2021
[23]

Cosdh: Communication-efficient collaborative perception via supply- demand awareness and intermediate-late hybridization

Junhao Xu, Yanan Zhang, Zhi Cai, and Di Huang. Cosdh: Communication-efficient collaborative perception via supply- demand awareness and intermediate-late hybridization. In Proceedings of the Computer Vision and Pattern Recognition Conference (CVPR), pages 6834–6843, 2025. 2

2025
[24]

InEuropean Conference on Computer Vision, pages 181–194

RuoyuXuandZhiyuXiang.Rlnet: Adaptivefusionof4dradar and lidar for 3d object detection. InEuropean Conference on Computer Vision, pages 181–194. Springer, 2025. 2

2025
[25]

Opencda: an open cooperative driving automation framework integrated with co-simulation

Runsheng Xu, Yi Guo, Xu Han, Xin Xia, Hao Xiang, and Jiaqi Ma. Opencda: an open cooperative driving automation framework integrated with co-simulation. In 2021 IEEE International Intelligent Transportation Systems Conference (ITSC), pages 1155–1162. IEEE, 2021. 2

2021
[26]

Opv2v: An open benchmark dataset and fusion pipelineforperceptionwithvehicle-to-vehiclecommunication

Runsheng Xu, Hao Xiang, Xin Xia, Xu Han, Jinlong Li, and Jiaqi Ma. Opv2v: An open benchmark dataset and fusion pipelineforperceptionwithvehicle-to-vehiclecommunication. In2022InternationalConferenceonRoboticsandAutomation (ICRA), pages 2583–2589, 2022. 2, 5, 7

2022
[27]

Spatio-temporal domain awareness for multi-agent collaborative perception

Kun Yang, Dingkang Yang, Jingyu Zhang, Mingcheng Li, Yang Liu, Jing Liu, Hanqi Wang, Peng Sun, and Liang Song. Spatio-temporal domain awareness for multi-agent collaborative perception. InProceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 233...

2023
[28]

V2x-radar: A multi-modal dataset with 4d radar for cooperative perception.arXiv preprint arXiv:2411.10962, 2024

LeiYang,XinyuZhang,ChenWang,JunLi,JiaqiMa,Zhiying Song, Tong Zhao, Ziying Song, Li Wang, Mo Zhou, Yang Shen, and Chen Lv. V2x-radar: A multi-modal dataset with 4d radar for cooperative perception.arXiv preprint arXiv:2411.10962, 2024. 1, 8

2024
[29]

3DSSD: Point-Based3DSingleStageObjectDetector

Zetong Yang, Yanan Sun, Shu Liu, and Jiaya Jia. 3DSSD: Point-Based3DSingleStageObjectDetector. In CVPR,2020. 5

2020
[30]

Mar- ius Zöllner

Melih Yazgan, Mythra Varun Akkanapragada, and J. Mar- ius Zöllner. Collaborative perception datasets in autonomous driving: A survey. In2024 IEEE Intelligent Vehicles Sympo- sium (IV), pages 2269–2276, 2024. 2

2024
[31]

In 2024 IEEE Intelligent Vehicles Symposium (IV), pages 2226–2233, 2024

Melih Yazgan, Thomas Graf, Min Liu, Tobias Fleck, and J.MariusZöllner.Asurveyonintermediatefusionmethodsfor collaborative perception categorized by real world challenges. In 2024 IEEE Intelligent Vehicles Symposium (IV), pages 2226–2233, 2024. 1, 2

2024
[32]

Efficomm: Bandwidth efficient multi agent communication

Melih Yazgan, Allen Xavier Arasan, and J Marius Zöllner. Efficomm: Bandwidth efficient multi agent communication. arXiv preprint arXiv:2507.19354, 2025. 2

2025 arXiv
[33]

Dair-v2x: A large-scale dataset for vehicle- infrastructurecooperative3dobjectdetection

Haibao Yu, Yizhen Luo, Mao Shu, Yiyi Huo, Zebang Yang, Yifeng Shi, Zhenglong Guo, Hanyu Li, Xing Hu, Jirui Yuan, and Zaiqing Nie. Dair-v2x: A large-scale dataset for vehicle- infrastructurecooperative3dobjectdetection. In Proceedings oftheIEEE/CVFConferenceonComputerVisionandP...

2022
[34]

Streamlts: Query-based temporal-spatial lidar fusion for cooperative object detection

Yunshuang Yuan and Monika Sester. Streamlts: Query-based temporal-spatial lidar fusion for cooperative object detection. InThe European Conference on Computer Vision Workshop (ECCVW), 2024. 2

2024
[35]

Bm2cp: Efficient collaborative perception with lidar-camera modalities.arXiv preprint arXiv:2310.14702, 2023

BinyuZhao,WeiZhang,andZhaonianZou. Bm2cp: Efficient collaborative perception with lidar-camera modalities.arXiv preprint arXiv:2310.14702, 2023. 6

2023 arXiv
[36]

Deformabledetr: Deformabletransformersfor end-to-endobjectdetection

Xizhou Zhu, Weĳie Su, Lewei Lu, Bin Li, Xiaogang Wang, andJifengDai. Deformabledetr: Deformabletransformersfor end-to-endobjectdetection. arXivpreprintarXiv:2010.04159 ,

2010 arXiv

Pith tools

Reviewed August 15, 2026 · model on record in the stance chip above.

[1] [1]

Ahmed, Siegfried Mercelis, and Ali Anwar

Ahmed N. Ahmed, Siegfried Mercelis, and Ali Anwar. Delawarecol: Delay aware collaborative perception.IEEE OpenJournalofVehicularTechnology ,6:1164–1177,2025. 2

work page 2025

[2] [2]

IgalBilik.Comparativeanalysisofradarandlidartechnologies for automotive applications.IEEE Intelligent Transportation Systems Magazine, 15(1):244–269, 2023. 1

work page 2023

[3] [3]

Cooper: Cooperative perception for connected autonomous vehicles based on 3d point clouds

Qi Chen, Sihai Tang, Qing Yang, and Song Fu. Cooper: Cooperative perception for connected autonomous vehicles based on 3d point clouds. In2019 IEEE 39th International ConferenceonDistributedComputingSystems(ICDCS) ,pages 514–524, 2019. 2

work page 2019

[4] [4]

CARLA: An open urban driving simulator

Alexey Dosovitskiy, German Ros, Felipe Codevilla, Antonio Lopez, and Vladlen Koltun. CARLA: An open urban driving simulator. InProceedings of the 1st Annual Conference on Robot Learning, pages 1–16, 2017. 2

work page 2017

[5] [5]

Deep residual learning for image recognition.arXiv preprint arXiv:1512.03385, 2015

Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition.arXiv preprint arXiv:1512.03385, 2015. 3

arXiv 2015

[6] [6]

Where2comm: Communication-efficient collaborative perceptionviaspatialconfidencemaps

Yue Hu, Shaoheng Fang, Zixing Lei, Yiqi Zhong, and Siheng Chen. Where2comm: Communication-efficient collaborative perceptionviaspatialconfidencemaps. In AdvancesinNeural Information Processing Systems, 2022. 2, 5, 6, 7

work page 2022

[7] [7]

V2x-r: Coop- erative lidar-4d radar fusion with denoising diffusion for 3d object detection

XunHuang,JinlongWang,QimingXia,SihengChen,Bisheng Yang, Xin Li, Cheng Wang, and Chenglu Wen. V2x-r: Coop- erative lidar-4d radar fusion with denoising diffusion for 3d object detection. InProceedings of the Computer Vision and PatternRecognitionConference(CVPR) ,pages27390–27400,

work page

[8] [8]

Adver-City: Open-Source Multi-Modal Dataset for Collaborative Perception Under Ad- verse Weather Conditions, 2025

Mateus Karvat and Sidney Givigi. Adver-City: Open-Source Multi-Modal Dataset for Collaborative Perception Under Ad- verse Weather Conditions, 2025. arXiv:2410.06380 [cs]. 2, 6

arXiv 2025