Pith. sign in

REVIEW 3 major objections 5 minor 45 references

SComCP: Task-Oriented Semantic Communication for Collaborative Perception

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 By transmitting only a task-selected subset of LiDAR semantic features through a deep joint source–channel codec, SComCP keeps 3D detection accuracy within 1.3% of the lossless upper bound at 0 dB SNR and a compression ratio of…

desk verdict A competent incremental JSCC-for-perception paper whose headline compression-ratio claim is undercut by a missing accounting of how the receiver knows which features were selected. read the letter →

arxiv 2507.00895 v1 pith:C3NG52VN submitted 2025-07-01 eess.SP

Jipeng Gan , Yucheng Sheng , Hua Zhang , Le Liang , Hao Ye , Chongtao Guo , Shi Jin This is my paper
classification eess.SP
keywords semanticcommunicationcollaborativeperceptionjointsource-channelcodingimportance-awarefeatureselection3DobjectdetectionV2VLiDARpointcloudsOPV2Vdataset
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

This paper proposes SComCP, an end-to-end framework for vehicle-to-vehicle collaborative 3D object detection in which a collaborating vehicle transmits only a tiny, task-relevant subset of its LiDAR-derived semantic features. The aim is to show that a joint source–channel coding codec plus an importance-aware selection network can keep perception accuracy close to the ideal all-features upper bound even under noisy channels. The paper reports that at a compression ratio of $1.4\times 10^{-3}$ and 0 dB SNR, its AP@0.5 is only 1.3% below the upper bound and 4.0% better than the baseline scheme, with the AP@0.7 gain widening to 8.8%. A sympathetic reader would care because this addresses the practical bottleneck that makes collaborative perception unsafe or bandwidth-prohibitive in real wireless conditions.

What carries the argument

The load-bearing object is the pair formed by the importance-aware feature selection network and the deep JSCC semantic codec. The selection network computes a spatial importance map, refines it with a residual cross-attention over the semantic feature map and a spatial attention module, softmax-normalizes the product, and thresholds it at $\gamma$ to produce a binary mask; this decides which $K$ feature vectors are kept and makes the compression ratio adaptive to scene content. The codec's encoder alternates transformer modules (long-range dependencies) with channel-attention modules (per-channel scaling and offset) and outputs power-normalized complex symbols $\mathbf{Z}_j$; the decoder mirrors the encoder and is trained jointly with the channel as a non-trainable layer, so source coding, channel coding, and modulation are learned as one operation rather than as separate stages.

What would settle it

Run the same OPV2V evaluation with the collaborator's projected point cloud perturbed by translation errors of 0.1–0.5 m and rotation errors of 1–5 degrees before feature extraction, and compare AP@0.7 at 0 dB SNR; if the margin over the baseline narrows to near zero or reverses, the claim that SComCP preserves accuracy under adverse wireless conditions depends on perfect alignment rather than on the codec alone.

Watch

Extended reading notes

Core claim

The central claim is that by jointly optimizing what to send and how to protect it, a collaborative perception system can transmit roughly one in seven hundred of its semantic feature values and still detect objects nearly as well as if it had sent everything over an ideal channel. The selection network learns a binary mask from an importance map refined by cross-attention and spatial attention, and a threshold $\gamma$ controls how many features survive, so the transmission volume adapts to scene richness. The codec, built from alternating transformer and channel-attention modules, maps selected features directly to power-normalized complex channel symbols, and its mirror-image decoder reconstructs them; because it is trained end to end under a Rayleigh fading channel, it avoids the cliff effect of separate source–channel coding. On the OPV2V benchmark the framework holds its advantage across AWGN and Rayleigh channels despite being trained on only one channel model.

Load-bearing premise

The results assume the collaborating vehicles' poses are known exactly and their transmissions synchronized, so the projected point clouds have no spatial misalignment; if pose error or clock skew is realistic, the measured detection gains may not hold.

Editorial extensions

If this is right

  • At a compression ratio of $1.4\times 10^{-3}$ and 0 dB SNR, SComCP's AP@0.5 sits only 1.3% below the lossless upper bound while using a small fraction of the bandwidth.
  • The framework beats the importance-map baseline by 4.0% at AP@0.5 and by 8.8% at AP@0.7 at 0 dB, with the gap widening as the channel worsens.
  • Unlike LDPC plus QAM schemes, SComCP shows a gradual performance decline rather than a cliff effect at low SNR.
  • A model trained only on Rayleigh fading carries its performance to AWGN channels without retraining, which the paper attributes to the codec's learned representation.
  • Raising the selection threshold $\gamma$ lowers the compression ratio while keeping AP@0.7 nearly flat across the tested CR values, so the framework can trade bandwidth against accuracy at deployment time.

Reading between the lines

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

  • If the selection network's importance map is reliable, the same mask could drive transmit-power or subcarrier allocation per feature, turning a binary selection into a graded resource-allocation signal.
  • The paper's performance numbers all assume perfect pose alignment and synchronized transmissions; a natural stress test is to add realistic pose noise and clock offsets, under which the reported gains may shrink.
  • The variable-length output controlled by $\gamma$ suggests a direct interface for rate control in a live V2V scheduler, where the available channel uses vary per frame.
  • Because the codec trains under one channel model and generalizes to another, a similar scheme may transfer to other modalities such as camera or radar features with the same task-oriented loss, although the paper does not test this.
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 SComCP, an end-to-end task-oriented semantic communication framework for V2V collaborative perception. A collaborator extracts a BEV semantic feature map, learns a binary mask that selects task-critical spatial locations, encodes the selected features with a deep JSCC codec, transmits the resulting channel symbols over a noisy channel, and the ego vehicle decodes, fuses with its own features, and performs 3D object detection. The main empirical claims are that at CR = 1.4e-3 and SNR = 0 dB the method is only 1.3% below the ideal upper bound in AP@0.5 and improves over a baseline by 4.0% in AP@0.5 and 8.8% in AP@0.7, while avoiding the cliff effect of separated LDPC-based schemes. Experiments are on OPV2V under AWGN and Rayleigh channels, with ablations for the selection network and the semantic codec.

Significance. If the central claims hold, SComCP is a useful practical design: it combines content-adaptive spatial feature selection with learned joint source-channel coding, achieving graceful degradation at low SNR with very low compression ratios on a standard collaborative-perception benchmark. The use of a public dataset (OPV2V), a standard PointPillars backbone, and an explicit equal-channel-use comparison against LDPC/16QAM and LDPC/256QAM are strengths. The ablations separating the contributions of the selection network and the codec are also valuable and clearly presented. However, the headline compression-ratio claim is not yet established because the paper does not specify how the receiver learns the selected spatial locations or account for the cost of that side information, and the empirical support would be stronger with error bars, multiple seeds, and an independent baseline beyond the same-group method [35].

major comments (3)
  1. [III-B, Eqs. (3)-(8); IV-B, Eq. (24)] The receiver is never told which spatial locations were selected. The binary mask Ω is computed from M_j and applied before transmission, but the reconstructed features are denoted \(\hat{F}_j \in \mathbb{R}^{K\times C}\u005d), a compact tensor with no spatial coordinates, and the fusion stage \(\Phi_{\mathrm{fus}}(M_e, \hat{F}_j)\u005d) in Eq. (8) must place these features on the ego BEV grid. The text in Sec. III-C says that zero-padding at the decoder output restores features to their original dimensions, which requires the mask or the selected indices at the receiver. If Ω is transmitted as side information, that cost is omitted from the channel-use budget in Eq. (24), so the reported CR = 1.4×10^{-3} is understated and the equal-channel-use comparison with the LDPC baselines is unfair; if Ω is not transmitted, Eq. (8) is not implementable as written. The authors must specify how the receiver obtains Ω and include its cost in the CR/channel-use accounting, or redesign the selection so that positions are recoverable without side information.
  2. [IV-C] The headline gains (1.3% below the upper bound, 4.0% and 8.8% over the baseline at 0 dB) are reported as single runs without error bars, confidence intervals, or multiple seeds. Given the small margins involved, please report means and variances over at least three independent training runs, and add at least one independent semantic-communication baseline beyond the same-group method [35] to support the general superiority claim.
  3. [II-A] The system model assumes “accurate vehicle poses and synchronized message transmissions” and states that this projection “ensures that subsequent feature fusion is free from spatial misalignment errors.” Since real V2V collaboration includes pose estimation error and clock skew, and the selection and fusion networks are trained under this ideal-assumption pipeline, the paper should either relax this assumption or explicitly evaluate sensitivity to pose and synchronization errors (e.g., by perturbing the OPV2V ground-truth poses during evaluation).
minor comments (5)
  1. [III-D and IV-D, Eqs. (15), (22), Fig. 10] The symbol γ is used for two different quantities: the selection threshold in Eq. (15) and the MSE loss weight in Eq. (22). Fig. 10 refers to “different values of γ” for the threshold, which is confusing; please use distinct symbols (e.g., τ for the threshold and λ for the loss weight).
  2. [IV-B] The sentence “Based on (25), the CR values ...” refers to Eq. (24), not Eq. (25); there is no Eq. (25) in the manuscript.
  3. [IV-D, Fig. 9] The text refers to “blue and yellow curves,” but the figure has no legend labels; please add a legend or explicitly identify the curves in the caption.
  4. [IV-D, Fig. 10] The horizontal axis of Fig. 10 is difficult to read: the relationship between the top row of numbers (1.5, 1.4, 1.34, ...) and the bottom row (10^{-3}, 10^{-4}) is unclear; please label the axis properly and state which quantity is varied.
  5. [III-A] The framework overview says the fusion network “projects the collaborator’s point cloud into the ego vehicle’s coordinate frame,” but the fusion actually operates on features, not raw point clouds; please rephrase to avoid confusion.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the central claims are empirical measurements against the external OPV2V benchmark, and the same-group citations are used only as benchmarks or related work.

full rationale

The paper contains no analytic derivation that could reduce to its inputs. The load-bearing results in Sec. IV-C (e.g., 1.3% gap to the upper bound at CR=1.4e-3 and SNR=0 dB, and 4.0%/8.8% gains over the baseline) are measured AP values on the OPV2V test set after end-to-end training of the selection network, JSCC codec, fusion, and detection network. The selection mask in Eqs. (3)-(4) and the codec in Eqs. (5)-(7) are trained against detection and MSE losses, so the reported performance is a fitted empirical outcome, not a quantity forced by construction or by the definition of the objective. The same-group baseline [35] is used only as a comparison scheme in Sec. IV-A and as related work; it is not cited to justify the design, to define the objective, or to exclude alternatives, so it is not load-bearing self-citation. No uniqueness theorem or ansatz is imported from the authors' prior work; architectural components such as PointPillars [38], Where2Comm [24], and CBAM [43] are external citations. The concern that the receiver is not told which spatial indices were selected is an overhead-accounting and implementation-specification issue, not a circularity in the paper's derivation chain.

Assumptions & free parameters 2 free parameters · 4 assumptions · 0 invented entities

This ledger counts the assumptions the empirical claim rests on. The system assumes ideal pose synchronization, a Gaussian and Rayleigh channel model, and simulator fidelity; the selection threshold is a hand-tuned free parameter and the MSE loss weight is unreported. No new physical entities are introduced.

free parameters (2)
  • Selection threshold gamma (Eq 15) = Not explicitly reported; average CR=1.4e-3 used in main experiments; Fig 10 sweeps threshold values 0.2, 0.5, 0.8…
    Controls the binary mask Omega and hence the compression ratio and communication load; the central trade-off claim depends on this choice, yet the exact default is not given.
  • MSE loss weight gamma (Eq 22) = Not reported
    Weights the transmission reconstruction loss against the detection loss in stage 2; no sensitivity analysis is provided.
assumptions (4)
  • domain assumption Accurate vehicle poses and synchronized message transmissions; projection into the ego frame yields zero spatial misalignment
    Introduced in Sec II-A; fusion and detection results are conditional on no pose or timing errors, which is untested.
  • domain assumption V2V wireless channel is modeled by multiplicative gain plus AWGN (Eq 6), with Rayleigh fading at training
    Sec II-A and Sec IV-A; the claimed robustness and generalization only cover this model family.
  • domain assumption OPV2V (CARLA/SUMO) is a representative proxy for real LiDAR collaborative perception
    Sec IV-A; no real-world data or field test is used.
  • domain assumption Learned JSCC codec can be optimized end-to-end with detection and MSE losses to provide robust transmission
    Sec III-C and III-D; the paper assumes deep JSCC approximates good source-channel coding without a theoretical guarantee.

how reviews work

0 comments
Cite this review

Pith. "Pith review of SComCP: Task-Oriented Semantic Communication for Collaborative Perception." pith.science (2026). https://pith.science/paper/C3NG52VN

@misc{pith2026250700895,
  author       = {Pith},
  title        = {Pith review of: SComCP: Task-Oriented Semantic Communication for Collaborative Perception},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/C3NG52VN}},
  note         = {Machine review of arXiv:2507.00895}
}
read the original abstract

Reliable detection of surrounding objects is critical for the safe operation of connected automated vehicles (CAVs). However, inherent limitations such as the restricted perception range and occlusion effects compromise the reliability of single-vehicle perception systems in complex traffic environments. Collaborative perception has emerged as a promising approach by fusing sensor data from surrounding CAVs with diverse viewpoints, thereby improving environmental awareness. Although collaborative perception holds great promise, its performance is bottlenecked by wireless communication constraints, as unreliable and bandwidth-limited channels hinder the transmission of sensor data necessary for real-time perception. To address these challenges, this paper proposes SComCP, a novel task-oriented semantic communication framework for collaborative perception. Specifically, SComCP integrates an importance-aware feature selection network that selects and transmits semantic features most relevant to the perception task, significantly reducing communication overhead without sacrificing accuracy. Furthermore, we design a semantic codec network based on a joint source and channel coding (JSCC) architecture, which enables bidirectional transformation between semantic features and noise-tolerant channel symbols, thereby ensuring stable perception under adverse wireless conditions. Extensive experiments demonstrate the effectiveness of the proposed framework. In particular, compared to existing approaches, SComCP can maintain superior perception performance across various channel conditions, especially in low signal-to-noise ratio (SNR) scenarios. In addition, SComCP exhibits strong generalization capability, enabling the framework to maintain high performance across diverse channel conditions, even when trained with a specific channel model.

Figures

Figures reproduced from arXiv: 2507.00895 by the authors.

Figure 1
Figure 1. An illustrative example of collaborative perceptio [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. An illustration of the proposed SComCP framework. [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗
Figure 3
Figure 3. Network structure of our proposed importance-aware [PITH_FULL_IMAGE:figures/full_fig_p005_3.png] view at source ↗
Figures from the paper (8 more)
Figure 4
Figure 4. Figure 4: Network structure of our proposed semantic codec net [PITH_FULL_IMAGE:figures/full_fig_p006_4.png]
Figure 5
Figure 5. Figure 5: Network structure of our proposed CA module. [PITH_FULL_IMAGE:figures/full_fig_p007_5.png]
Figure 6
Figure 6. Figure 6: Performance of our proposed method compared with oth [PITH_FULL_IMAGE:figures/full_fig_p009_6.png]
Figure 7
Figure 7. Figure 7: Performance of our proposed method compared with oth [PITH_FULL_IMAGE:figures/full_fig_p010_7.png]
Figure 8
Figure 8. Figure 8: Examples of 3D detection outputs produced by our prop [PITH_FULL_IMAGE:figures/full_fig_p011_8.png]
Figure 10
Figure 10. Figure 10: Performance of our proposed method with different t [PITH_FULL_IMAGE:figures/full_fig_p011_10.png]
Figure 11
Figure 11. Figure 11: Performance of our proposed method compared with ot [PITH_FULL_IMAGE:figures/full_fig_p012_11.png]
Figure 12
Figure 12. Figure 12: Performance of our proposed method compared with ot [PITH_FULL_IMAGE:figures/full_fig_p012_12.png]

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

45 extracted references · 43 canonical work pages

  1. [35]

    Semantic c ommunica- tion for cooperative perception based on importance map,

    Y . Sheng, H. Y e, L. Liang, S. Jin, and G. Y . Li, “Semantic c ommunica- tion for cooperative perception based on importance map,” J. Franklin Inst., vol. 361, no. 6, p. 106739, Mar. 2024

    work page 2024

  2. [1]

    6G for vehicle-to-ever ything (V2X) communications: Enabling technologies, challenges , and oppor- tunities,

    M. Noor-A-Rahim, Z. Liu, H. Lee, M. O. Khyam, J. He, D. Pesc h, K. Moessner, W. Saad, and H. V . Poor, “6G for vehicle-to-ever ything (V2X) communications: Enabling technologies, challenges , and oppor- tunities,” Proc. IEEE , vol. 110, no. 6, pp. 712–734, May 2022. 13

    work page 2022

  3. [2]

    Level-5 autonomous driving—Are we there yet ? A review of research literature,

    M. A. Khan, H. E. Sayed, S. Malik, T. Zia, J. Khan, N. Alkaab i, and H. Ignatious, “Level-5 autonomous driving—Are we there yet ? A review of research literature,” ACM Comput. Surv., vol. 55, no. 2, pp. 1–38, Jan. 2022

    work page 2022

  4. [3]

    Towards the unified pri nciples for level 5 autonomous vehicles,

    J. Wang, H. Huang, K. Li, and J. Li, “Towards the unified pri nciples for level 5 autonomous vehicles,” Eng., vol. 7, no. 9, pp. 1313–1325, Sept. 2021

    work page 2021

  5. [4]

    Multi-modal 3D object det ection in autonomous driving: A survey and taxonomy,

    L. Wang, X. Zhang, Z. Song, J. Bi, G. Zhang, H. Wei, L. Tang, L. Y ang, J. Li, C. Jia, and L. Zhao, “Multi-modal 3D object det ection in autonomous driving: A survey and taxonomy,” IEEE Trans. Intell. V ehicles, vol. 8, no. 7, pp. 3781–3798, Apr. 2023

    work page 2023

  6. [5]

    Robustness-aware 3D object detection in autonomous drivi ng: A review and outlook,

    Z. Song, L. Liu, F. Jia, Y . Luo, C. Jia, G. Zhang, L. Y ang, an d L. Wang, “Robustness-aware 3D object detection in autonomous drivi ng: A review and outlook,” IEEE Trans. Intell. Transp. Syst. , Aug. 2024

    work page 2024

  7. [6]

    Cooperative perception wit h V2V communication for autonomous vehicles,

    H. Ngo, H. Fang, and H. Wang, “Cooperative perception wit h V2V communication for autonomous vehicles,” IEEE Trans. V eh. Technol. , vol. 72, no. 9, pp. 11 122–11 131, Sept. 2023

    work page 2023

  8. [7]

    Collabo rative perception in autonomous driving: Methods, datasets, and c hallenges,

    Y . Han, H. Zhang, H. Li, Y . Jin, C. Lang, and Y . Li, “Collabo rative perception in autonomous driving: Methods, datasets, and c hallenges,” IEEE Intell. Transp. Syst. Mag. , vol. 15, no. 6, pp. 131–151, Sept. 2023

    work page 2023

Show all 45 references
  1. [8]

    Towards vehicle-to-everything au- tonomous driving: A survey on collaborative perception,

    S. Liu, C. Gao, Y . Chen, X. Peng, X. Kong, K. Wang, R. Xu, W. Jiang, H. Xiang, J. Ma et al. , “Towards vehicle-to-everything au- tonomous driving: A survey on collaborative perception,” arXiv preprint arXiv:2308.16714, 2023

    2023 arXiv

  2. [9]

    V2I interference mana gement for collaborative perception based on hierarchical reinforce ment learning,

    G. Liu, H. Y e, L. Liang, and S. Jin, “V2I interference mana gement for collaborative perception based on hierarchical reinforce ment learning,” in Proc. IEEE 34th Int. Conf. Mach. Learn. Signal Process. (MLS P), Sept. 2024, pp. 1–6

    2024

  3. [10]

    Spectrum sharing in vehicu lar networks based on multi-agent reinforcement learning,

    L. Liang, H. Y e, and G. Y . Li, “Spectrum sharing in vehicu lar networks based on multi-agent reinforcement learning,” IEEE J. Sel. Areas Com- mun., vol. 37, no. 10, pp. 2282–2292, 2019

    2019

  4. [11]

    Bringing different views together: A hybrid cooperative perception framework for connected autonomous vehicles,

    D. Carrillo, M. Nutt, M. Meijer, J. Khan, S. Fu, and Q. Y an g, “Bringing different views together: A hybrid cooperative perception framework for connected autonomous vehicles,” IEEE Netw., pp. 1–1, Feb. 2025

    2025

  5. [12]

    Cooper: Cooperativ e perception for connected autonomous vehicles based on 3D point clouds,

    Q. Chen, S. Tang, Q. Y ang, and S. Fu, “Cooper: Cooperativ e perception for connected autonomous vehicles based on 3D point clouds, ” in Proc. IEEE 39th Int. Conf. Distrib. Comput. Syst. (ICDCS) , Jul. 2019, pp. 514–524

    2019

  6. [13]

    Virtual- painting: Addressing sparsity with virtual points and dist ance-aware data augmentation for 3D object detection,

    S. Dhakal, D. Carrillo, D. Qu, M. Nutt, Q. Y ang, and S. Fu, “Virtual- painting: Addressing sparsity with virtual points and dist ance-aware data augmentation for 3D object detection,” arXiv preprint arXiv:2312.16141, 2023

    2023 arXiv

  7. [14]

    Collaborative automated d riving: A machine learning-based method to enhance the accuracy of sh ared information,

    Z. Y . Rawashdeh and Z. Wang, “Collaborative automated d riving: A machine learning-based method to enhance the accuracy of sh ared information,” in Proc. 21st Int. Conf. Intell. Transp. Syst. (ITSC) , November 2018, pp. 3961–3966

    2018

  8. [15]

    CLOCs: Camera-lidar o bject candidates fusion for 3D object detection,

    S. Pang, D. Morris, and H. Radha, “CLOCs: Camera-lidar o bject candidates fusion for 3D object detection,” in Proc. IEEE/RSJ Int. Conf. Intell. Robots Syst. (IROS) , Oct. 2020, pp. 10 386–10 393

    2020

  9. [16]

    Asynchrony-robust collaborative perception via bird’s e ye view flow,

    S. Wei, Y . Wei, Y . Hu, Y . Lu, Y . Zhong, S. Chen, and Y . Zhang , “Asynchrony-robust collaborative perception via bird’s e ye view flow,” in Proc. Adv. Neural Inf. Process. Syst. (NeurIPS) , vol. 36, 2023, pp. 28 462–28 477

    2023

  10. [17]

    PACP: Priority-aware collaborative perception f or connected and autonomous vehicles,

    Z. Fang, S. Hu, H. An, Y . Zhang, J. Wang, H. Cao, X. Chen, an d Y . Fang, “PACP: Priority-aware collaborative perception f or connected and autonomous vehicles,” IEEE Trans. Mob. Comput. , vol. 23, no. 12, pp. 15 003–15 018, Aug. 2024

    2024

  11. [18]

    Practical collaborative perception: A framework for asyn chronous and multi-agent 3D object detection,

    M.-Q. Dao, J. S. Berrio, V . Fr´ emont, M. Shan, E. H´ ery, a nd S. Worrall, “Practical collaborative perception: A framework for asyn chronous and multi-agent 3D object detection,” IEEE Trans. Intell. Transp. Syst. , vol. 25, no. 9, pp. 12 163–12 175, Sept. 2024

    2024

  12. [19]

    A survey and tutorial of redundancy mitigation for vehicular cooperative percepti on: Standards, strategies and open issues,

    T. Lyu, D. Pesch, A. O’Driscoll et al. , “A survey and tutorial of redundancy mitigation for vehicular cooperative percepti on: Standards, strategies and open issues,” arXiv preprint arXiv:2501.01200 , 2025

    2025 arXiv

  13. [20]

    A survey of collaborative perception in intelligent vehicles at intersections,

    X. Gao, X. Zhang, Y . Lu, Y . Huang, L. Y ang, Y . Xiong, and P . Liu, “A survey of collaborative perception in intelligent vehicles at intersections,” IEEE Trans. Intell. V ehicles, pp. 1–20, May 2024

    2024

  14. [21]

    When2com: Mu lti-agent perception via communication graph grouping,

    Y .-C. Liu, J. Tian, N. Glaser, and Z. Kira, “When2com: Mu lti-agent perception via communication graph grouping,” in Proc. IEEE/CVF Conf. Comput. Vis. Pattern Recognit. (CVPR) , 2020, pp. 4106–4115

    2020

  15. [22]

    V2VNet: V ehicle-to-vehicle communication for joi nt perception and prediction,

    T.-H. Wang, S. Manivasagam, M. Liang, B. Y ang, W. Zeng, a nd R. Ur- tasun, “V2VNet: V ehicle-to-vehicle communication for joi nt perception and prediction,” in Proc. Eur . Conf. Comput. Vis. (ECCV) , 2020, pp. 605–621

    2020

  16. [23]

    Learn ing distilled collaboration graph for multi-agent perception,

    Y . Li, S. Ren, P . Wu, S. Chen, C. Feng, and W. Zhang, “Learn ing distilled collaboration graph for multi-agent perception,” Proc. Adv. Neural Inf. Process. Syst. (NeurIPS) , vol. 34, pp. 29 541–29 552, 2021

    2021

  17. [24]

    Where2comm : Communication-efficient collaborative perception via spa tial confidence maps,

    Y . Hu, S. Fang, Z. Lei, Y . Zhong, and S. Chen, “Where2comm : Communication-efficient collaborative perception via spa tial confidence maps,” Proc. Adv. Neural Inf. Process. Syst. (NeurIPS) , vol. 35, pp. 4874–4886, 2022

    2022

  18. [25]

    Communication chal- lenges in infrastructure-vehicle cooperative autonomous driving: A field deployment perspective,

    S. Liu, B. Y u, J. Tang, Y . Zhu, and X. Liu, “Communication chal- lenges in infrastructure-vehicle cooperative autonomous driving: A field deployment perspective,” IEEE Wireless Commun. , vol. 29, no. 4, pp. 126–131, Aug. 2022

    2022

  19. [26]

    Challenges and s olutions for cellular based V2X communications,

    S. Gyawali, S. Xu, Y . Qian, and R. Q. Hu, “Challenges and s olutions for cellular based V2X communications,” IEEE Commun. Surv. Tut. , vol. 23, no. 1, pp. 222–255, Oct. 2020

    2020

  20. [27]

    Joint coding and decisi on feedback equalization for broadband wireless channels,

    S. L. Ariyavisitakul and Y . Li, “Joint coding and decisi on feedback equalization for broadband wireless channels,” IEEE J. Sel. Areas Commun., vol. 16, no. 9, pp. 1670–1678, Dec. 1998

    1998

  21. [28]

    Semantic c om- munication for cooperative perception using HARQ,

    Y . Sheng, L. Liang, H. Y e, S. Jin, and G. Y . Li, “Semantic c om- munication for cooperative perception using HARQ,” arXiv preprint arXiv:2409.09042, 2024

    2024 arXiv

  22. [29]

    Deep re inforce- ment learning-based user scheduling for collaborative per ception,

    Y . Liu, G. Liu, L. Liang, H. Y e, C. Guo, and S. Jin, “Deep re inforce- ment learning-based user scheduling for collaborative per ception,” arXiv preprint arXiv:2502.10456, 2025

    2025 arXiv

  23. [30]

    AI empowered wireless communications: From bits to semantics ,

    Z. Qin, L. Liang, Z. Wang, S. Jin, X. Tao, W. Tong, and G. Y . Li, “AI empowered wireless communications: From bits to semantics ,” Proc. IEEE, vol. 112, no. 7, pp. 621–652, Jul. 2024

    2024

  24. [31]

    ISCom: Interest-aware semantic communication scheme for point cloud video streaming on metaverse XR devices,

    Y . Huang, B. Bai, Y . Zhu, X. Qiao, X. Su, L. Y ang, and P . Zha ng, “ISCom: Interest-aware semantic communication scheme for point cloud video streaming on metaverse XR devices,” IEEE J. Sel. Areas Commun., vol. 42, no. 4, pp. 1003–1021, Apr. 2023

    2023

  25. [32]

    T ask- oriented semantic communication for stereo-vision 3D obje ct detection,

    Z. Cao, H. Zhang, L. Liang, H. Wang, S. Jin, and G. Y . Li, “T ask- oriented semantic communication for stereo-vision 3D obje ct detection,” IEEE Trans. Commun. , pp. 1–1, 2025

    2025

  26. [33]

    Semantic communication systems for speech transmission,

    Z. Weng and Z. Qin, “Semantic communication systems for speech transmission,” IEEE J. Sel. Areas Commun. , vol. 39, no. 8, pp. 2434– 2444, Jun. 2021

    2021

  27. [34]

    Dee p joint source- channel coding for wireless image transmission,

    E. Bourtsoulatze, D. Burth Kurka, and D. G¨ und¨ uz, “Dee p joint source- channel coding for wireless image transmission,” IEEE Trans. Cogn. Commun. Netw., vol. 5, no. 3, pp. 567–579, May 2019

    2019

  28. [36]

    A comprehensive survey of lidar-based 3d object detection me thods with deep learning for autonomous driving,

    G. Zamanakos, L. Tsochatzidis, A. Amanatiadis, and I. P ratikakis, “A comprehensive survey of lidar-based 3d object detection me thods with deep learning for autonomous driving,” Comput. Graph. , vol. 99, pp. 153–181, Oct. 2021

    2021

  29. [37]

    Generalized intersection over union: A metric and a loss fo r bounding box regression,

    H. Rezatofighi, N. Tsoi, J. Gwak, A. Sadeghian, I. Reid, a nd S. Savarese, “Generalized intersection over union: A metric and a loss fo r bounding box regression,” in Proc. IEEE/CVF Conf. Comput. Vis. Pattern Recog- nit. (CVPR) , 2019, pp. 658–666

    2019

  30. [38]

    Pointpillars: Fast encoders for object detection from poi nt clouds,

    A. H. Lang, S. V ora, H. Caesar, L. Zhou, J. Y ang, and O. Bei jbom, “Pointpillars: Fast encoders for object detection from poi nt clouds,” in Proc. IEEE/CVF Conf. Comput. Vis. Pattern Recognit. (CVPR) , 2019, pp. 12 697–12 705

    2019

  31. [39]

    V2X- ViT: V ehicle-to-everything cooperative perception with visio n transformer,

    R. Xu, H. Xiang, Z. Tu, X. Xia, M.-H. Y ang, and J. Ma, “V2X- ViT: V ehicle-to-everything cooperative perception with visio n transformer,” in Proc. Eur . Conf. Comput. Vis. (ECCV) , 2022, pp. 107–124

    2022

  32. [40]

    OPV2V: An open benchmark dataset and fusion pipeline for perception with v ehicle-to- vehicle communication,

    R. Xu, H. Xiang, X. Xia, X. Han, J. Li, and J. Ma, “OPV2V: An open benchmark dataset and fusion pipeline for perception with v ehicle-to- vehicle communication,” in Proc. Int. Conf. Robotics Autom. (ICRA) , 2022, pp. 2583–2589

    2022

  33. [41]

    Task-oriented communicat ion for vehicle- to-infrastructure cooperative perception,

    J. Shao, T. Li, and J. Zhang, “Task-oriented communicat ion for vehicle- to-infrastructure cooperative perception,” in Proc. IEEE 34th Int. Conf. Mach. Learn. Signal Process. (MLSP) , Sept. 2024, pp. 1–6

    2024

  34. [42]

    Attention is all you need,

    A. V aswani, N. Shazeer, N. Parmar, J. Uszkoreit, L. Jone s, A. N. Gomez, Ł. Kaiser, and I. Polosukhin, “Attention is all you need,” Proc. Adv. Neural Inf. Process. Syst. (NeurIPS) , vol. 30, 2017

    2017

  35. [43]

    CBAM: Convol utional block attention module,

    S. Woo, J. Park, J.-Y . Lee, and I. S. Kweon, “CBAM: Convol utional block attention module,” in Proc. Eur . Conf. Comput. Vis. (ECCV), 2018, pp. 3–19

    2018

  36. [44]

    Focal loss for dense object detection,

    T.-Y . Lin, P . Goyal, R. Girshick, K. He, and P . Doll´ ar, “ Focal loss for dense object detection,” in Proc. IEEE Int. Conf. Comput. Vision (ICCV) , 2017, pp. 2980–2988

    2017

  37. [45]

    CARLA: An open urban driving simulator,

    A. Dosovitskiy, G. Ros, F. Codevilla, A. Lopez, and V . Ko ltun, “CARLA: An open urban driving simulator,” in Proc. Annu. Conf. Robot Learn. (CoRL) , 2017, pp. 1–16

    2017

Pith tools

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