Pith. sign in

REVIEW 3 major objections 5 minor 47 references

CoST: Efficient Collaborative Perception From Unified Spatiotemporal Perspective

T0 review · 3 major / 5 minor · reviewed 2026-08-06 · deepseek-v4-flash

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

Pith's one-line read V2X detection cuts bandwidth by reusing static scenes once.

desk verdict Solid empirical gains but the efficiency claim is contradicted by the paper's own selection mask; rewrite needed. read the letter →

arxiv 2508.00359 v1 pith:HLXTDB2M submitted 2025-08-01 cs.CV

Zongheng Tang , Yi Liu , Yifan Sun , Yulu Gao , Jinyu Chen , Runsheng Xu , Si Liu This is my paper
classification cs.CV
keywords collaborativeperceptionspatio-temporalfusionV2Xcommunication3DobjectdetectionbandwidthreductionBEVfeaturesdeformableattentionLiDARpointclouds
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

CoST, a framework for collaborative perception in autonomous driving, claims that fusing space and time into a single spatio-temporal space—instead of doing multi-agent fusion and temporal fusion as separate steps—improves both accuracy and communication efficiency. The paper's key move is to treat historical agents as time-delayed duplicates of current agents, so static objects appear only once in the unified space and their features need be transmitted only once, while dynamic objects are transmitted selectively. The practical stakes are real: on the V2V4Real, V2XSet, and DAIR-V2X benchmarks, CoST reports higher 3D detection average precision than prior methods while cutting transmission volume by up to two orders of magnitude, and its modules can be attached to existing fusion methods.

What carries the argument

The load-bearing object is the unified spatio-temporal space, in which a historical agent is treated literally as a current agent with a fixed delay, collapsing spatial and temporal fusion into one operation. Two mechanisms carry the bandwidth claim: the STT selection mask $M_i^t = E_i^t \cdot (\frac{1}{\rho+1} + D_i^t \cdot \rho)$, which blends a saliency map $E_i^t$ with the absolute temporal difference $D_i^t = |E_i^t - E_i^{t-\tau}|$ to decide which tokens are dynamic enough to transmit, and a reconstruction module that merges the transmitted sparse tokens with pose-warped historical features recovered from a memory bank. The fusion claim is carried by recurrent USTF, which keeps only one pose-projected historical feature $B^{t-\tau}$ and feeds it into the same multi-agent fusion as an extra delayed agent, and by MADA, a deformable-attention fusion engine whose Align Transformation injects per-agent metadata such as velocity and latency.

What would settle it

A decisive test is to run CoST on a scene where everything moves (no static objects at all) and check whether the bandwidth saving from STT disappears; a second test is to inject pose errors of about one meter into the warping and measure whether detection AP falls below the best baseline, which would show that reused static features are being corrupted by alignment drift.

Watch

Extended reading notes

Core claim

The central discovery is that static object features are redundant across time in collaborative perception: once historical agents are pose-warped onto the current frame to define a shared spatio-temporal space, each static object yields a single observation that can be stored and reused instead of re-transmitted. CoST operationalizes this with a Spatio-temporal Transmission module that computes a classification saliency map and its temporal difference, blends them into a selection mask, and transmits only the selected dynamic tokens; a lightweight reconstruction network then combines those tokens with the pose-aligned historical features from a memory bank to recover the full feature map. A Unified Spatio-temporal Fusion module treats the fused historical feature as one additional agent with a fixed time delay, so multi-agent and temporal fusion run as a single operation implemented by a Multi-Agent Deformable Attention that injects velocity, latency, and infrastructure metadata. The reported results show CoST beating the compared methods on all three datasets at both AP@0.5 and AP@0.7 while transmitting far fewer feature tokens.

Load-bearing premise

The claimed bandwidth savings rest on the selection mask correctly separating static from moving content and on the pose-based warping of history being accurate enough that reused static features are not stale.

Editorial extensions

If this is right

  • Attaching STT to existing fusion methods such as V2X-ViT and CoBEVT reduces their transmission bandwidth while slightly raising AP, so the efficiency gain does not depend on CoST's own fusion design.
  • Because USTF retains a single fused historical feature instead of a window of past frames, the unified spatio-temporal fusion adds only a small computational overhead over single-frame fusion.
  • The latency and dropout experiments in the paper imply that the reconstructed static features act as a memory that softens the effect of missing or delayed transmissions, a direct corollary of transmitting each static object once.
  • Thresholding the selection mask provides a single control knob that trades bandwidth against accuracy across a wide range of compression rates, letting one trained model adapt to different channel budgets.

Reading between the lines

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

  • If the static/dynamic split is reliable, the same reuse idea could extend to other persistent representations such as occupancy grids or HD maps, where the bandwidth saving would be even larger.
  • The fixed-delay treatment of historical agents suggests a natural generalization to asynchronous agents with heterogeneous latencies, where each agent is tagged with its own delay and fused in the same shared space.
  • The temporal-difference-driven mask is a learned rate-control signal, so a natural extension is to train the mask with an explicit bandwidth penalty, making the trade-off adaptive to channel conditions instead of set by a fixed threshold.
  • Because the reconstructed static features come from a pose warp, a testable extension is to share a single reconstruction decoder across all agents, which would cut both computation and memory further.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

3 major / 5 minor

Summary. The paper proposes CoST, a collaborative perception framework that unifies spatial (multi-agent) and temporal fusion by treating historical agents as time-delayed copies of current agents. It introduces three components: Spatio-temporal Transmission (STT), which selects features for transmission to reduce bandwidth; Unified Spatio-temporal Fusion (USTF), which fuses current and historical agents in a single module; and Multi-Agent Deformable Attention (MADA), an efficient cross-agent attention mechanism. Experiments on V2V4Real, V2XSet, and DAIR-V2X show consistent accuracy gains over prior methods, and integration with V2X-ViT and CoBEVT demonstrates that the proposed modules can be plugged into existing frameworks. The central claims are that CoST improves detection accuracy while reducing transmission bandwidth, and that the bandwidth reduction comes from transmitting dynamic content only once while reusing static features from memory.

Significance. If the claims hold, CoST would be a practical contribution to V2X collaborative perception, offering both accuracy and efficiency improvements. The unified spatio-temporal perspective is conceptually appealing, and the paper provides a broad set of experiments across three datasets, component ablations, robustness to latency and pose noise, and plug-in results on two existing methods. The proposed STT and USTF modules are presented as framework-agnostic, which would increase their utility. However, the efficiency claim is currently under-specified and internally inconsistent as written, and the experimental reporting lacks variance information. These issues need to be resolved before the contribution can be fully assessed.

major comments (3)
  1. [Section 3.2, Algorithm 1] The selection mask M_i^t is defined as a real-valued map, and z_i^t = M_i^t ⊙ F_i^t is computed without any thresholding or sparsification step. As written, z_i^t is dense and the communication volume is identical to transmitting the full feature map. The bandwidth reduction claimed in Section 4.5 depends on 'thresholding M_i^t', but this operation is not described in Algorithm 1 or in the training procedure. Please specify the exact binarization rule, how the threshold is applied during training (e.g., straight-through estimator, soft mask, or hard mask with detached gradients), and how the number of transmitted elements N_c in Eq. (6) is computed from the thresholded mask.
  2. [Section 3.2, Eq. (5) and Section 4.5] The claim that static objects are transmitted only once is not supported by the mask formula. For a static region, D_i^t ≈ 0, so M_i^t = E_i^t/(ρ+1). A high-saliency static object (E_i^t ≈ 1) then has M_i^t ≈ 1/(ρ+1), which exceeds the typical threshold of 0.01 whenever ρ < 99. The paper never reports the value of ρ, so the reader cannot determine whether high-saliency static objects are retransmitted every frame. Consequently, the 60× compression attributed to static-feature reuse in Figure 5 is not justified. Please report ρ, the threshold used in the main experiments, an analysis of the fraction of transmitted tokens that correspond to dynamic versus static regions, and a sensitivity study over these hyperparameters.
  3. [Section 4.2 and Table 1] All reported results are single-run numbers without error bars or multiple-seed variance. On DAIR-V2X, the AP@0.7 gain over Where2comm is only 0.013 (0.473 vs. 0.460), which may be within run-to-run noise. Please provide means and standard deviations over at least three independent training runs for the main results, or otherwise justify that the observed gains are statistically significant.
minor comments (5)
  1. [Eq. (6)] The communication volume formula is ambiguous: the divisor '8 × 220' should likely be '8 × 2^20' with a clear unit conversion to megabytes. Please clarify the expression and state whether N_c counts the number of selected BEV tokens or the number of transmitted feature elements.
  2. [Figure 10 caption] The caption refers to 'CoTT' in the text; this appears to be a typo for 'CoST'.
  3. [Section 4.2] The sentence 'We compare our results with [18]' is confusing because reference [18] is a survey rather than a list of baseline methods. Please write out the compared methods explicitly in the text.
  4. [Abstract / Code availability] The abstract states 'Code will be available', but the repository is not yet public, which limits reproducibility. If possible, release the code or provide a detailed supplementary description of implementation choices.
  5. [Table 3] The row labels and the text in Section 4.4 are slightly inconsistent: the text says 'Increasing the interval to 2 provides the highest performance, with an AP@0.5 of 70.96%', but Table 3 shows 0.7064 for two input frames at interval 2 and 0.7096 for three input frames at interval 2. Please clarify which configuration achieves which number.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: CoST's accuracy and bandwidth claims are benchmarked on public datasets against published baselines, and the STT efficiency claim, while internally under-specified, is not derived from its own inputs by construction.

full rationale

The paper's central accuracy claim is supported by evaluations on V2V4Real, V2XSet, and DAIR-V2X against published baselines, with no parameter fitted to the test set being renamed as a prediction. The STT module (Algorithm 1) computes a mask from a saliency map and a temporal-difference map; the resulting compression ratio is measured from the transmitted token count, and reconstruction quality is measured by downstream detection AP on held-out data. The abstract's statement that 'each static object yields a single observation... and thus only requires transmission only once' is an interpretive summary of the design rather than a quantity derived from the equations. Indeed, the equations do not force static-region masks to zero: when D_t^i is near zero, M_t^i = E_t^i / (rho+1), so high-saliency static objects can still be selected unless rho and the threshold are set appropriately. That is a potential internal-consistency and missing-detail concern about the 'once-only' claim, but it is not a circular reduction, because the mask is not defined in terms of the claimed outcome and the empirical bandwidth-accuracy trade-off is evaluated externally. Self-citations to datasets and methods by co-author Runsheng Xu (e.g., V2X-ViT, OPV2V, CoBEVT, V2V4Real) are used as benchmarks and comparison points, not as load-bearing justification for CoST's own claims; no uniqueness theorem or prior result is invoked to force the design. The MADA module borrows deformable attention from external prior work and is tested against strong baselines. Overall, the derivation chain is self-contained with respect to external benchmarks, so no circular step is identifiable.

Assumptions & free parameters 3 free parameters · 5 assumptions · 0 invented entities

The central claims rest on benchmark data quality, the static-scene assumption behind STT, the sufficiency of saliency-difference for detecting motion, and the recurrent-history approximation in USTF. No new physical entities are postulated.

free parameters (3)
  • rescale rate rho = not reported
    Blends saliency and dynamic maps in the selection mask in Algorithm 1 line 5; no value or tuning procedure is given in the paper.
  • transmission threshold on M_i^t = 0.01 for about 60x compression
    Section 4.5 says varying the threshold in [0.0001, 1] controls the bandwidth-accuracy trade-off; the operating point is chosen by hand.
  • time interval tau during training = 2 frames
    Supplementary Section 8 sets tau to 2 frames, and the ablation in Table 3 shows AP varies with tau, so the result depends on this choice.
assumptions (5)
  • domain assumption Public benchmark annotations and sensor calibration in V2V4Real, V2XSet, and DAIR-V2X are correct and comparable across methods.
    The central empirical claim depends on these datasets, described in Section 4.1.
  • domain assumption Static scene content is unchanged between t and t-tau, so pose-warped historical features can substitute for current static features in reconstruction.
    STT reconstructs R_i^t from sparse tokens plus R_i^{t-tau} in Section 3.2 and Algorithm 1; if static areas change or pose alignment is imperfect, reconstructed features are stale.
  • domain assumption Classification-head saliency and its temporal difference identify dynamic objects well enough to drive the selection mask.
    M_i^t is built from E_i^t and D_i^t in Eqs. (1)-(5); no ground-truth motion supervision is used for this mask.
  • domain assumption A single recurrent collaborative feature B^{t-tau} projected to time t is an adequate substitute for all per-agent historical features.
    USTF feeds one delayed agent into the fusion module instead of a parallel window of all agents, as described in Section 3.3.
  • domain assumption Deformable attention remains effective when sampling points are drawn from multiple agents' BEV features.
    MADA in Eq. (5) generalizes multi-scale deformable attention; no proof is given that the multi-agent extension preserves the original convergence or accuracy properties.

how reviews work

0 comments
Cite this review

Pith. "Pith review of CoST: Efficient Collaborative Perception From Unified Spatiotemporal Perspective." pith.science (2026). https://pith.science/paper/HLXTDB2M

@misc{pith2026250800359,
  author       = {Pith},
  title        = {Pith review of: CoST: Efficient Collaborative Perception From Unified Spatiotemporal Perspective},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/HLXTDB2M}},
  note         = {Machine review of arXiv:2508.00359}
}
read the original abstract

Collaborative perception shares information among different agents and helps solving problems that individual agents may face, e.g., occlusions and small sensing range. Prior methods usually separate the multi-agent fusion and multi-time fusion into two consecutive steps. In contrast, this paper proposes an efficient collaborative perception that aggregates the observations from different agents (space) and different times into a unified spatio-temporal space simultanesouly. The unified spatio-temporal space brings two benefits, i.e., efficient feature transmission and superior feature fusion. 1) Efficient feature transmission: each static object yields a single observation in the spatial temporal space, and thus only requires transmission only once (whereas prior methods re-transmit all the object features multiple times). 2) superior feature fusion: merging the multi-agent and multi-time fusion into a unified spatial-temporal aggregation enables a more holistic perspective, thereby enhancing perception performance in challenging scenarios. Consequently, our Collaborative perception with Spatio-temporal Transformer (CoST) gains improvement in both efficiency and accuracy. Notably, CoST is not tied to any specific method and is compatible with a majority of previous methods, enhancing their accuracy while reducing the transmission bandwidth.

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

47 extracted references · 40 canonical work pages

  1. [1]

    Cooperative perception for 3d object detec- tion in driving scenarios using infrastructure sensors

    Eduardo Arnold, Mehrdad Dianati, Robert De Temple, and Saber Fallah. Cooperative perception for 3d object detec- tion in driving scenarios using infrastructure sensors. IEEE Transactions on Intelligent Transportation Systems , 23(3): 1852–1864, 2020. 2

    work page 2020

  2. [2]

    End-to- end object detection with transformers

    Nicolas Carion, Francisco Massa, Gabriel Synnaeve, Nicolas Usunier, Alexander Kirillov, and Sergey Zagoruyko. End-to- end object detection with transformers. In European confer- ence on computer vision, pages 213–229. Springer, 2020. 3

    2020

  3. [3]

    F-cooper: Feature based cooperative perception for autonomous vehicle edge computing system using 3d point clouds

    Qi Chen, Xu Ma, Sihai Tang, Jingda Guo, Qing Yang, and Song Fu. F-cooper: Feature based cooperative perception for autonomous vehicle edge computing system using 3d point clouds. In Proceedings of the 4th ACM/IEEE Symposium on Edge Computing, pages 88–100, 2019. 6

    work page 2019

  4. [4]

    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. In 2019 IEEE 39th International Conference on Distributed Computing Systems (ICDCS) , pages 514–524. IEEE, 2019. 2

    work page 2019

  5. [5]

    Transiff: An instance-level feature fusion framework for vehicle- infrastructure cooperative 3d detection with transformers

    Ziming Chen, Yifeng Shi, and Jinrang Jia. Transiff: An instance-level feature fusion framework for vehicle- infrastructure cooperative 3d detection with transformers. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 18205–18214, 2023. 6

    work page 2023

  6. [6]

    Carla: An open urban driv- ing simulator

    Alexey Dosovitskiy, German Ros, Felipe Codevilla, Anto- nio Lopez, and Vladlen Koltun. Carla: An open urban driv- ing simulator. In Conference on robot learning, pages 1–16. PMLR, 2017. 12

    work page 2017

  7. [7]

    Multi- agent collaborative perception via motion-aware robust com- munication network

    Shixin Hong, Yu Liu, Zhi Li, Shaohui Li, and You He. Multi- agent collaborative perception via motion-aware robust com- munication network. In Proceedings of the IEEE/CVF Con- ference on Computer Vision and Pattern Recognition, pages 15301–15310, 2024. 3

    work page 2024

  8. [8]

    Where2comm: Communication-efficient collab- orative perception via spatial confidence maps

    Yue Hu, Shaoheng Fang, Zixing Lei, Yiqi Zhong, and Si- heng Chen. Where2comm: Communication-efficient collab- orative perception via spatial confidence maps. Advances in neural information processing systems , 35:4874–4886,

    work page

Show all 47 references
  1. [9]

    Bevdet4d: Exploit tempo- ral cues in multi-camera 3d object detection

    Junjie Huang and Guan Huang. Bevdet4d: Exploit tempo- ral cues in multi-camera 3d object detection. arXiv preprint arXiv:/2203.17054, 2021. 3

    2021 arXiv

  2. [10]

    Adam: A method for stochastic optimization

    Diederik P Kingma and Jimmy Ba. Adam: A method for stochastic optimization. arXiv preprint arXiv:1412.6980 ,

    arXiv

  3. [11]

    Pointpillars: Fast encoders for object detection from point clouds

    Alex H Lang, Sourabh V ora, Holger Caesar, Lubing Zhou, Jiong Yang, and Oscar Beijbom. Pointpillars: Fast encoders for object detection from point clouds. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 12697–12705, 2019. 3, 6, 7, 11, 12

    2019

  4. [12]

    Latency-aware collaborative perception

    Zixing Lei, Shunli Ren, Yue Hu, Wenjun Zhang, and Siheng Chen. Latency-aware collaborative perception. In European Conference on Computer Vision , pages 316–332. Springer,

  5. [13]

    Learning distilled collaboration graph for multi-agent perception

    Yiming Li, Shunli Ren, Pengxiang Wu, Siheng Chen, Chen Feng, and Wenjun Zhang. Learning distilled collaboration graph for multi-agent perception. Advances in Neural Infor- mation Processing Systems, 34:29541–29552, 2021. 2

    2021

  6. [14]

    Bevdepth: Acquisition of reliable depth for multi-view 3d object detec- tion

    Yinhao Li, Zheng Ge, Guanyi Yu, Jinrong Yang, Zengran Wang, Yukang Shi, Jianjian Sun, and Zeming Li. Bevdepth: Acquisition of reliable depth for multi-view 3d object detec- tion. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 1477–1485, 2023. 3

    2023

  7. [15]

    Bevformer: Learning bird’s-eye-view representation from multi-camera images via spatiotemporal transformers

    Zhiqi Li, Wenhai Wang, Hongyang Li, Enze Xie, Chong- hao Sima, Tong Lu, Yu Qiao, and Jifeng Dai. Bevformer: Learning bird’s-eye-view representation from multi-camera images via spatiotemporal transformers. In European con- ference on computer vision, pages 1–18. Springer, 2022. 3

    2022

  8. [16]

    Focal loss for dense object detection

    Tsung-Yi Lin, Priya Goyal, Ross Girshick, Kaiming He, and Piotr Doll´ar. Focal loss for dense object detection. In Pro- ceedings of the IEEE international conference on computer vision, pages 2980–2988, 2017. 4, 11

    2017

  9. [17]

    Sparse4d: Multi-view 3d object detec- tion with sparse spatial-temporal fusion

    Xuewu Lin, Tianwei Lin, Zixiang Pei, Lichao Huang, and Zhizhong Su. Sparse4d: Multi-view 3d object detec- tion with sparse spatial-temporal fusion. arXiv preprint arXiv:2211.10581, 2022. 3

    2022 arXiv

  10. [18]

    Towards vehicle-to-everything autonomous driving: A survey on collaborative perception.arXiv preprint arXiv:2308.16714, 2023

    Si Liu, Chen Gao, Yuan Chen, Xingyu Peng, Xianghao Kong, Kun Wang, Runsheng Xu, Wentao Jiang, Hao Xiang, Jiaqi Ma, et al. Towards vehicle-to-everything autonomous driving: A survey on collaborative perception.arXiv preprint arXiv:2308.16714, 2023. 6

    2023 arXiv

  11. [19]

    Select2col: Leveraging spatial-temporal impor- tance of semantic information for efficient collaborative per- ception

    Yuntao Liu, Qian Huang, Rongpeng Li, Xianfu Chen, Zhifeng Zhao, Shuyuan Zhao, Yongdong Zhu, and Hong- gang Zhang. Select2col: Leveraging spatial-temporal impor- tance of semantic information for efficient collaborative per- ception. IEEE Transactions on Vehicular Technology, 2024. 3

    2024

  12. [20]

    Who2com: Collaborative perception via learnable handshake communication

    Yen-Cheng Liu, Junjiao Tian, Chih-Yao Ma, Nathan Glaser, Chia-Wen Kuo, and Zsolt Kira. Who2com: Collaborative perception via learnable handshake communication. In 2020 IEEE International Conference on Robotics and Automation (ICRA), pages 6876–6883. IEEE, 2020. 2, 3

    2020

  13. [21]

    Decoupled weight decay regularization

    Ilya Loshchilov and Frank Hutter. Decoupled weight decay regularization. arXiv preprint arXiv:1711.05101, 2017. 12

    2017 arXiv

  14. [22]

    Disconet: Shapes learning on dis- connected manifolds for 3d editing

    Eloi Mehr, Ariane Jourdan, Nicolas Thome, Matthieu Cord, and Vincent Guitteny. Disconet: Shapes learning on dis- connected manifolds for 3d editing. In Proceedings of the IEEE/CVF International Conference on Computer Vision , pages 3474–3483, 2019. 6

    2019

  15. [23]

    Time will tell: New outlooks and a baseline for temporal multi- view 3d object detection

    Jinhyung Park, Chenfeng Xu, Shijia Yang, Kurt Keutzer, Kris M Kitani, Masayoshi Tomizuka, and Wei Zhan. Time will tell: New outlooks and a baseline for temporal multi- view 3d object detection. InThe Eleventh International Con- ference on Learning Representations, 2022. 3

    2022

  16. [24]

    Faster r-cnn: Towards real-time object detection with region proposal networks

    Shaoqing Ren, Kaiming He, Ross Girshick, and Jian Sun. Faster r-cnn: Towards real-time object detection with region proposal networks. Advances in neural information process- ing systems, 28, 2015. 4, 11

    2015

  17. [25]

    Vips: Real-time perception fusion for infrastructure-assisted au- tonomous driving

    Shuyao Shi, Jiahe Cui, Zhehao Jiang, Zhenyu Yan, Guo- liang Xing, Jianwei Niu, and Zhenchao Ouyang. Vips: Real-time perception fusion for infrastructure-assisted au- tonomous driving. In Proceedings of the 28th annual inter- national conference on mobile computing and networki...

    2022

  18. [26]

    Exploring object-centric temporal modeling for efficient multi-view 3d object detection

    Shihao Wang, Yingfei Liu, Tiancai Wang, Ying Li, and Xi- angyu Zhang. Exploring object-centric temporal modeling for efficient multi-view 3d object detection. In Proceedings of the IEEE/CVF International Conference on Computer Vi- sion, pages 3621–3631, 2023. 3

    2023

  19. [27]

    Umc: A unified bandwidth-efficient and multi-resolution based collaborative perception framework

    Tianhang Wang, Guang Chen, Kai Chen, Zhengfa Liu, Bo Zhang, Alois Knoll, and Changjun Jiang. Umc: A unified bandwidth-efficient and multi-resolution based collaborative perception framework. In Proceedings of the IEEE/CVF In- ternational Conference on Computer Vision (ICCV) , ...

    2023

  20. [28]

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

    Tsun-Hsuan Wang, Sivabalan Manivasagam, Ming Liang, Bin Yang, Wenyuan Zeng, and Raquel Urtasun. V2vnet: Vehicle-to-vehicle communication for joint perception and prediction. In European conference on computer vision , pages 605–621. Springer, 2020. 1, 2, 6

    2020

  21. [29]

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

    Runsheng Xu, Yi Guo, Xu Han, Xin Xia, Hao Xiang, and Jiaqi Ma. Opencda: an open cooperative driving automa- tion framework integrated with co-simulation. In 2021 IEEE International Intelligent Transportation Systems Conference (ITSC), pages 1155–1162. IEEE, 2021. 12

    2021

  22. [30]

    Bridging the domain gap for multi-agent perception

    Runsheng Xu, Jinlong Li, Xiaoyu Dong, Hongkai Yu, and Ji- aqi Ma. Bridging the domain gap for multi-agent perception. arXiv preprint arXiv:2210.08451, 2022. 2

    2022 arXiv

  23. [31]

    V2x-vit: Vehicle-to-everything cooperative perception with vision transformer

    Runsheng Xu, Hao Xiang, Zhengzhong Tu, Xin Xia, Ming- Hsuan Yang, and Jiaqi Ma. V2x-vit: Vehicle-to-everything cooperative perception with vision transformer. InEuropean conference on computer vision , pages 107–124. Springer,

  24. [32]

    Opv2v: An open benchmark dataset and fusion pipeline for perception with vehicle-to-vehicle communica- tion

    Runsheng Xu, Hao Xiang, Xin Xia, Xu Han, Jinlong Li, and Jiaqi Ma. Opv2v: An open benchmark dataset and fusion pipeline for perception with vehicle-to-vehicle communica- tion. In 2022 International Conference on Robotics and Au- tomation (ICRA), pages 2583–2589. IEEE, 2022. 1,...

    2022

  25. [33]

    Cobevt: Cooperative bird’s eye view semantic segmentation with sparse transformers

    Runsheng Xu, Zhengzhong Tu, Hao Xiang, Wei Shao, Bolei Zhou, and Jiaqi Ma. Cobevt: Cooperative bird’s eye view semantic segmentation with sparse transformers. In Confer- ence on Robot Learning, pages 989–1000. PMLR, 2023. 6, 7, 8

    2023

  26. [34]

    V2v4real: A real-world large-scale dataset for vehicle-to-vehicle cooperative perception

    Runsheng Xu, Xin Xia, Jinlong Li, Hanzhao Li, Shuo Zhang, Zhengzhong Tu, Zonglin Meng, Hao Xiang, Xiaoyu Dong, Rui Song, et al. V2v4real: A real-world large-scale dataset for vehicle-to-vehicle cooperative perception. In Proceed- ings of the IEEE/CVF conference on computer vis...

    2023

  27. [35]

    How2comm: Communication-efficient and collaboration-pragmatic multi- agent perception

    Dingkang Yang, Kun Yang, Yuzheng Wang, Jing Liu, Zhi Xu, Rongbin Yin, Peng Zhai, and Lihua Zhang. How2comm: Communication-efficient and collaboration-pragmatic multi- agent perception. Advances in Neural Information Process- ing Systems, 36:25151–25164, 2023. 1, 3

    2023

  28. [36]

    What2comm: Towards communication-efficient collaborative perception via feature decoupling

    Kun Yang, Dingkang Yang, Jingyu Zhang, Hanqi Wang, Peng Sun, and Liang Song. What2comm: Towards communication-efficient collaborative perception via feature decoupling. In Proceedings of the 31st ACM international conference on multimedia, pages 7686–7695, 2023. 6

    2023

  29. [37]

    Dair-v2x: A large-scale dataset for vehicle- infrastructure cooperative 3d object detection

    Haibao Yu, Yizhen Luo, Mao Shu, Yiyi Huo, Zebang Yang, Yifeng Shi, Zhenglong Guo, Hanyu Li, Xing Hu, Jirui Yuan, et al. Dair-v2x: A large-scale dataset for vehicle- infrastructure cooperative 3d object detection. In Proceed- ings of the IEEE/CVF conference on computer vision a...

    2022

  30. [38]

    Flow-based feature fusion for vehicle- infrastructure cooperative 3d object detection

    Haibao Yu, Yingjuan Tang, Enze Xie, Jilei Mao, Ping Luo, and Zaiqing Nie. Flow-based feature fusion for vehicle- infrastructure cooperative 3d object detection. Advances in Neural Information Processing Systems, 36, 2024. 1, 3

    2024

  31. [39]

    Dsdnet: Deep structured self- driving network

    Wenyuan Zeng, Shenlong Wang, Renjie Liao, Yun Chen, Bin Yang, and Raquel Urtasun. Dsdnet: Deep structured self- driving network. In European conference on computer vi- sion, pages 156–172. Springer, 2020. 2

    2020

  32. [40]

    Deformable detr: Deformable trans- formers for end-to-end object detection

    Xizhou Zhu, Weijie Su, Lewei Lu, Bin Li, Xiaogang Wang, and Jifeng Dai. Deformable detr: Deformable trans- formers for end-to-end object detection. arXiv preprint arXiv:2010.04159, 2020. 5 CoST: Efficient Collaborative Perception From Unified Spatiotemporal Perspective Supplem...

    2010 arXiv

  33. [42]

    Problem Formulation In a collaborative perception scenario involving N agents, each agent i owns unique observations{Xi}N i=1 and the per- ception supervision {Yi}N i=1

    Preliminary 6.1. Problem Formulation In a collaborative perception scenario involving N agents, each agent i owns unique observations{Xi}N i=1 and the per- ception supervision {Yi}N i=1. The objective is to maximize the collective perception performance of all agents while ens...

  34. [43]

    Conventional methods reduce feature channels [31, 34], but they neglect the temporal redundancy where static informa- tion remains unchanged over time

    Spatiotemporal Transmission As shown in Figure 8, the STT module in our CoST frame- work reduces communication bandwidth by transmitting only dynamic object features instead of the full BEV map. Conventional methods reduce feature channels [31, 34], but they neglect the tempor...

  35. [44]

    Training Details All detection models utilize PointPillar [11] to extract the BEV features from the point cloud and the feature channel C is set as 256. The models are trained over 60 epochs with a batch size of 4 per GPU (Tesla V100), a learning rate set at 0.001, and incorpo...

  36. [45]

    Covering a driving area of 410 km, V2V4Real includes 20K LiDAR frames, 240K annotated 3D bounding boxes across 5 classes, and HDMaps that encompass all driving routes

    Datasets V2V4Real: V2VReal is a large-scale real-world V2V dataset collected by two vehicles with multi-modal sensors navigating through diverse scenarios. Covering a driving area of 410 km, V2V4Real includes 20K LiDAR frames, 240K annotated 3D bounding boxes across 5 classes,...

  37. [46]

    Our model produces precise bound- ing boxes that closely align with the actual ground truth, contrasting with competing methods that exhibit notable discrepancies

    Qualitative Analysis Figure 10 illustrates the detection visualization for OPV2V , CoBEVT, V2X-ViT, and CoTT across sparse and dense scenes in V2V4Real. Our model produces precise bound- ing boxes that closely align with the actual ground truth, contrasting with competing meth...

  38. [47]

    It is essential for collabo- rative perception techniques to effectively handle such time delays

    Robustness Analysis Time Delay poses a notable challenge in real-world V2X communications, leading to a lack of synchronization be- tween the ego vehicle’s functionalities and information re- ceived from collaborative agents. It is essential for collabo- rative perception tech...

  39. [2022]

    2, 4, 5, 6, 7, 8, 11, 12

Pith tools

Reviewed August 6, 2026 · model on record in the stance chip above.