Pith. sign in

REVIEW 3 major objections 4 minor 46 references

A wireless link that sends LiDAR point clouds as digital learned features reconstructs them more faithfully than classic compression plus channel coding, at half the data.

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

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

LPC-FT transmits LiDAR point clouds as compressed learned features over digital channels and, in OpV2V experiments, reduces Chamfer Distance by 30% and raises PSNR by 1.9 dB versus the SEPT baseline.

T0 review reviewed 2026-08-04 challenge →

load-bearing objection Plausible digital semantic-communication system for LiDAR point clouds, but the headline gains depend on an unaccounted side channel for point coordinates, plus missing code and thin baselines. the 3 major comments →

arxiv 2509.06506 v1 pith:TWDQ4RVC submitted 2025-09-08 eess.SP

Synesthesia of Machines (SoM)-Aided LiDAR Point Cloud Transmission for Collaborative Perception

Ensong Liu , Rongqing Zhang , Xiang Cheng , Jian Tang This is my paper
classification eess.SP
keywords LiDAR point cloud transmissioncollaborative perceptionsemantic communicationsjoint source-channel codingdensity-preserving point cloud compressionself-attention channel codingdigital modulation with straight-through estimatorOpV2V vehicle-to-vehicle dataset
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

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 tries to show that LiDAR point clouds, which are far too large for low-latency collaborative perception, can be transmitted as compact learned features over real digital wireless links without losing the structure needed for reconstruction. The proposed system, LPC-FT, combines a density-preserving point cloud autoencoder with self-attention channel coding, receiver-side cross-attention fusion, and a nonlinear straight-through estimator that makes digital modulation trainable. On the OpV2V collaborative-driving dataset, the authors report a 30% reduction in Chamfer Distance and a 1.9 dB improvement in point-to-plane PSNR over the best prior learned point cloud transmission method, using half the data. If this holds, it suggests that end-to-end learned digital semantic communication can replace separate octree compression and channel coding for LiDAR sharing between vehicles.

Core claim

The central claim is that a digital semantic communication system can transmit LiDAR point clouds for collaborative perception more faithfully than both traditional octree-based compression followed by LDPC channel coding and existing deep learning-based point cloud transmission methods. LPC-FT encodes the whole point cloud into a compact downsampled representation using a density-preserving compression network, adds an absolute coordinate position encoding to the point transformer blocks, and then protects the features with a self-attention channel encoder that also ingests a channel SNR embedding. At the receiver, a symmetric channel decoder cleans the noisy features, a cross-attention mod

What carries the argument

The central mechanism is the density-preserving point cloud autoencoder, which represents a LiDAR cloud as a downsampled point set Ps and a compact feature set Fs, decoupled from the channel code. The transmitted coordinates Ps are assumed error-free via a pre-shared anchor point set, while the features Fs are passed through a self-attention channel encoder with SNR embedding, quantized, modulated, sent over AWGN or Rayleigh fading, demodulated, and decoded. At the receiver, a cross-attention feature fusion module exploits the correlation between the transmitter's and receiver's point clouds to enhance reconstruction. The digital link is made trainable by a straight-through estimator paired

Load-bearing premise

The receiver must already have the exact downsampled point coordinates Ps before transmission; the paper assumes these arrive error-free through a pre-shared anchor point set, but if those coordinates are corrupted, delayed, or unavailable in a real wireless link, the entire reconstruction pipeline has no anchor on which to rebuild the point cloud.

What would settle it

Take the trained LPC-FT and transmit the anchor coordinates Ps through a noisy bitstream at the same SNR instead of assuming they are error-free, then measure reconstruction Chamfer Distance and PSNR; a sharp degradation with even small coordinate errors would show that the reported robustness depends on a side channel that the real wireless system does not provide.

Watch this falsifier. Get emailed when new claim-graph text bears on it.

If this is right

  • If the reported gains reproduce, LiDAR sharing between vehicles can operate at half the transmission data of prior learned methods while producing point clouds closer to the original, enabling lower-latency collaborative perception.
  • The system's digital modulation and demodulation stages make it directly compatible with existing digital communication hardware, unlike discrete-time analog transmission schemes that dominate semantic communication research.
  • The SNR embedding allows a single trained model to operate over a range of channel conditions, so vehicles do not need to retrain for each channel state.
  • The two-stage training strategy, with a pretrained compression model fine-tuned for the channel, means adapting to new channel conditions requires only a short fine-tuning run rather than full retraining.
  • Because the system reconstructs the complete environment rather than only object-level features, it can support downstream planning and prediction tasks that need background context, not just detection.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • A natural stress test would be to transmit the anchor coordinates Ps through the same wireless link with a modest coding overhead; if reconstructing the downsampled coordinates degrades gracefully with channel quality, the system's robustness claim would extend to a fully wireless setup, whereas any error in Ps would likely collapse reconstruction quality.
  • The cross-attention fusion assumes the transmitter and receiver point clouds are roughly aligned; a test with imperfect or missing extrinsic calibration between vehicles would reveal how much of the reported gain depends on accurate spatial registration.
  • The same bottleneck features could be reused for downstream object detection or trajectory prediction without full point cloud reconstruction, potentially making the semantic link serve multiple perception tasks with a single transmission.
  • The SNR-adaptive model could be extended to time-varying channels by feeding real-time channel estimates into the same SNR embedding, moving the system from block-fading-static assumptions toward more realistic dynamic vehicular channels.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

3 major / 4 minor

Summary. The paper proposes LPC-FT, a digital semantic communication system for LiDAR point cloud transmission in collaborative perception. The transmitter compresses a LiDAR point cloud with a density-preserving autoencoder, enhances the bottleneck features with a self-attention channel encoder, quantizes and modulates them, and the receiver applies a channel decoder, a cross-attention feature fusion module that uses the receiver's own point cloud, and a density-aware upsampling decoder to reconstruct the transmitter's cloud. Digital quantization/modulation is handled with a nonlinear straight-through estimator and tanh activation, and transfer learning is used to adapt to different SNRs. Experiments on OpV2V report that LPC-FT outperforms SEPT, an internal no-channel-coding ablation (LPC-FC), and octree+LDPC, with an average 30% Chamfer Distance reduction and 1.9 dB PSNR improvement over SEPT, at lower bpp.

Significance. If the results hold, the paper makes a useful contribution to semantic communications for LiDAR point clouds: it is one of the few works to consider a digital (quantized, modulated) transmission framework rather than discrete-time analog transmission, and the ablation study on the channel encoder and feature fusion modules is informative. The use of SNR embeddings, transfer learning, and the nonlinear STE is practical and well motivated. However, the central empirical claim is weakened by three issues: the load-bearing assumption that downsampled coordinates are delivered error-free with no protocol or bit cost, the absence of the underlying compression backbone [32] as a baseline, and the lack of error bars or significance measures. The paper is not reproducible as no code is released.

major comments (3)
  1. [Section III and Eq. (7)-(18)] The assumption that the downsampled coordinates Ps are delivered error-free via a pre-obtained 'anchor point set' is load-bearing and unquantified. The receiver conditions all reconstruction on Ps: Eq. (12) centers every upsampled point at Ps, Eqs. (15)-(18) use Ps in attention position encodings, and Eq. (7) decodes features aligned to Ps. If Ps is the FPS output of the current transmitter frame, the receiver cannot have it before transmission; if the anchor set is pre-shared and scene-independent, it is not the encoder's Ps. No protocol, bit budget, or protection mechanism is given, and no coordinate bit cost is included in the reported bpp. Because the headline gains are measured against baselines that also receive some free bits (first 1/8 for octree), the 'same data cost' claim is not established. Please specify how Ps is obtained/transmitted and include its cost in the comparison,
  2. [Section V-B and Figs. 7-8] The baseline set does not include the underlying compression backbone [32] that the feature encoder is explicitly adopted from. LPC-FC is an internal ablation, and SEPT is modified with an anchor-point strategy. Without a noiseless and digital-channel comparison to [32] at matched bpp, the reader cannot separate the contribution of the proposed channel/fusion/activation modules from the choice of the compression backbone. Please add [32] (or its re-implementation) at the same bpp and same coordinate side-information accounting, and report the noiseless performance of LPC-FT as an upper bound.
  3. [Section V-D] No error bars or significance tests are reported. The central numerical claims (30% CD reduction, 1.9 dB PSNR improvement) are averaged over test frames, but single-run curves in Figs. 7-8 cannot support confidence in these differences. Moreover, the text concedes that octree+LDPC with 1/2 rate outperforms LPC-FT on the point-to-point CD at SNR=5 dB, and the counter-argument depends on the factor-of-two data rate. With Ps side-information unaccounted, this comparison is not yet conclusive. Report variance/per-scene results and a rate-matched comparison.
minor comments (4)
  1. [Section IV-C, Eq. (18)] In the cross-attention equation, the key index appears to be a typo: eK_i should likely be eK_j, since the text describes attention between transmitter query i and receiver key j.
  2. [Throughout] There are numerous typos and grammatical errors, e.g., 'actiavtion', 'incorprate', 'transmssion', 'caculate', 'asssume'. A full proofread is needed.
  3. [Fig. 9] The ground-truth rows are labeled 'Bpp: 96' in every column, which is confusing; only reconstruction methods should carry bpp labels, and the ground-truth row should be clean.
  4. [Section III] The 'anchor point set' is mentioned but never formally defined. Please define it, distinguish it from the current frame's Ps, and specify whether it is pre-shared, updated, or transmitted.

Circularity Check

0 steps flagged

No circular derivation; the reported gains are empirical against external baselines. The error-free Ps side-channel assumption is a missing-support fairness risk, not a circularity.

full rationale

The central claims (30% CD reduction, 1.9 dB PSNR improvement) are established by simulations on OpV2V against external baselines (SEPT [40], octree+LDPC, and the LPC-FC ablation). No equation is defined in terms of the quantity it is supposed to predict, and no fitted parameter is renamed as a prediction. The only self-citations ([10] SoM framing, [19], [37] multi-modal semantic communications) are motivational and do not carry technical weight; no uniqueness theorem or ansatz is imported from them. The one load-bearing weakness is in Section III: "We assume the transmission of coordinates is error-free, because we can define an anchor point set that the transmitter and the receiver have already obtained before the transmission." This matters because Eq. (12) centers every upsampled neighborhood at points of Ps, and Eqs. (15)-(18) use Ps as the geometric conditioning of the attention modules. If Ps is the current frame's FPS output, the receiver cannot possess it before transmission; if the anchor set is pre-shared and scene-independent, it cannot be the encoder's Ps. No protocol is given for obtaining or protecting Ps, so the bitrate comparison may count coordinate delivery as free side information. This is a correctness/fairness risk, but it is not circular in the derivation sense: the reconstruction is still evaluated on held-out frames against external benchmarks, and the assumption does not make the output equal to an input by construction. Hence score 1 rather than 0 only to reflect that the self-citations, while non-load-bearing, are present.

Axiom & Free-Parameter Ledger

3 free parameters · 5 axioms · 1 invented entities

The central results rest on standard point-cloud and channel modeling assumptions plus a critical error-free coordinate transmission assumption. The only explicitly hand-set hyperparameters are the loss weights and bottleneck dimension; learned network weights are trained, not free parameters in this sense. The anchor point set is an invented protocol construct with no specified implementation.

free parameters (3)
  • density loss weight alpha = 5e-4
    Hand-set to weight the density term in Eq. 19, adopted from [32].
  • cardinality loss weight beta = 5e-6
    Hand-set to weight the cardinality term in Eq. 19, adopted from [32].
  • bottleneck feature dimension C = 8
    Author-chosen dimension of the transmitted feature per downsampled point; sets the communication rate.
axioms (5)
  • domain assumption Radio channel is block fading with known CSI; the coherence time equals the symbol block length NsC, and ZF equalization is applied.
    Section III, Eqs. 3-4: 'we assume the coherent time is NsC, which is also called block fading channel' and 'We assume channel h is known'.
  • domain assumption The downsampled point coordinates Ps are transmitted error-free via a pre-shared anchor point set.
    Section III: 'We assume the transmission of coordinates is error-free, because we can define an anchor point set that the transmitter and the receiver have already obtained before the transmission.'
  • domain assumption The receiver's own point cloud PR is available and is encoded with the same feature encoder S alpha to provide fusion features.
    Section III, Eq. 8: 'xR = S alpha(PR)'; the feature fusion module in Section IV-C depends on this.
  • domain assumption The density-preserving point cloud compression network of [32] is a valid autoencoder base, and its settings are reused.
    Section IV-A: 'The other network settings of the feature encoder are the same as the density-preserving point cloud compression network [32].'
  • domain assumption OpV2V simulated LiDAR frames are representative for evaluating the generalization of LPC-FT.
    Section V-A: all experiments use OpV2V; no real-world LiDAR test is provided.
invented entities (1)
  • Anchor point set no independent evidence
    purpose: Justifies the assumption that downsampled coordinates Ps need not be transmitted over the wireless channel, since transmitter and receiver 'have already obtained' them.
    Section III introduces this construct textually but gives no protocol, synchronization, or error-handling mechanism; it is a hand-waved enabler of the error-free coordinate assumption.

reviewed 2026-08-04 · how reviews work

0 comments
Cite this review

Pith. "Pith review of Synesthesia of Machines (SoM)-Aided LiDAR Point Cloud Transmission for Collaborative Perception." pith.science (2026). https://pith.science/paper/TWDQ4RVC

@misc{pith2026250906506,
  author       = {Pith},
  title        = {Pith review of: Synesthesia of Machines (SoM)-Aided LiDAR Point Cloud Transmission for Collaborative Perception},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/TWDQ4RVC}},
  note         = {Machine review of arXiv:2509.06506}
}
Share X Bluesky LinkedIn Reddit HN
read the original abstract

Collaborative perception enables more accurate and comprehensive scene understanding by learning how to share information between agents, with LiDAR point clouds providing essential precise spatial data. Due to the substantial data volume generated by LiDAR sensors, efficient point cloud transmission is essential for low-latency multi-agent collaboration. In this work, we propose an efficient, robust and applicable LiDAR point cloud transmission system via the Synesthesia of Machines (SoM), termed LiDAR Point Cloud Feature Transmission (LPC-FT), to support collaborative perception among multiple agents. Specifically, we employ a density-preserving deep point cloud compression method that encodes the complete point cloud into a downsampled efficient representation. To mitigate the effects of the wireless channel, we design a channel encoder module based on self-attention to enhance LiDAR point cloud features and a feature fusion module based on cross-attention to integrate features from transceivers. Furthermore, we utilize the nonlinear activation layer and transfer learning to improve the training of deep neural networks in the presence the digital channel noise. Experimental results demonstrate that the proposed LPC-FT is more robust and effective than traditional octree-based compression followed by channel coding, and outperforms state-of-the-art deep learning-based compression techniques and existing semantic communication methods, reducing the Chamfer Distance by 30% and improving the PSNR by 1.9 dB on average. Owing to its superior reconstruction performance and robustness against channel variations, LPC-FT is expected to support collaborative perception tasks.

Figures

Figures reproduced from arXiv: 2509.06506 by Ensong Liu, Jian Tang, Rongqing Zhang, Xiang Cheng.

Figure 1
Figure 1. Figure 1: Overview of the LPC-FT, our proposed LiDAR point clou [PITH_FULL_IMAGE:figures/full_fig_p003_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: The architecture of the feature encoder and decoder, [PITH_FULL_IMAGE:figures/full_fig_p005_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: The improved point transformer block with the absolu [PITH_FULL_IMAGE:figures/full_fig_p005_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: The channel encoder network for the LiDAR point cloud [PITH_FULL_IMAGE:figures/full_fig_p006_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: The feature fusion network for the transmitter and re [PITH_FULL_IMAGE:figures/full_fig_p007_5.png] view at source ↗
Figure 6
Figure 6. Figure 6: The digital communication system with the straight- [PITH_FULL_IMAGE:figures/full_fig_p008_6.png] view at source ↗
Figure 7
Figure 7. Figure 7: The reconstruction performance versus SNR under the [PITH_FULL_IMAGE:figures/full_fig_p010_7.png] view at source ↗
Figure 8
Figure 8. Figure 8: The reconstruction performance versus SNR under the [PITH_FULL_IMAGE:figures/full_fig_p010_8.png] view at source ↗
Figure 9
Figure 9. Figure 9: Qualitative results on OpV2V. From top to bottom: SNR [PITH_FULL_IMAGE:figures/full_fig_p011_9.png] view at source ↗
Figure 10
Figure 10. Figure 10: The reconstruction performance under various chan [PITH_FULL_IMAGE:figures/full_fig_p012_10.png] view at source ↗
Figure 11
Figure 11. Figure 11: Chamfer Distance loss versus training epochs for th [PITH_FULL_IMAGE:figures/full_fig_p013_11.png] view at source ↗

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

46 extracted references · 39 canonical work pages · 1 internal anchor

  1. [1]

    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, Nov.-Dec. 2023. JOURNAL OF LATEX CLASS FILES, VOL. 14, NO. 8, AUGUST 2021 14

    work page 2023

  2. [2]

    Lidar for autonomous drivin g: The principles, challenges, and trends for automotive lidar an d perception systems,

    Y . Li and J. Ibanez-Guzman, “Lidar for autonomous drivin g: The principles, challenges, and trends for automotive lidar an d perception systems,” IEEE Signal Process. Mag. , vol. 37, no. 4, pp. 50–61, Jul. 2020

    work page 2020

  3. [3]

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

    Q. Chen, S. Tang, Q. Y ang, and S. Fu, “Cooper: Cooperative perception for connected autonomous vehicles based on 3d point clouds, ” in Proc. IEEE Int. Conf. Distri. Comp. Syst. , Dallas, TX, USA, Jul. 2019, pp. 514–524

    work page 2019

  4. [4]

    Learning distilled collaboration graph for multi-agent perception,

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

    work page 2021

  5. [5]

    V2x-v it: 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-v it: V ehicle-to-everything cooperative perception with visio n transformer,” in Proc. Eur . Conf. Comput. Vis. , Berlin, Heidelberg, Oct. 2022, pp. 107–124

    work page 2022

  6. [6]

    Core: Coo perative reconstruction for multi-agent perception,

    B. Wang, L. Zhang, Z. Wang, Y . Zhao, and T. Zhou, “Core: Coo perative reconstruction for multi-agent perception,” in Proc. IEEE Int. Conf. Comput. Vis., Paris, France, Oct. 2023, pp. 8710–8720

    work page 2023

  7. [7]

    Learnin g for vehicle-to-vehicle cooperative perception under lossy co mmunication,

    J. Li, R. Xu, X. Liu, J. Ma, Z. Chi, J. Ma, and H. Y u, “Learnin g for vehicle-to-vehicle cooperative perception under lossy co mmunication,” IEEE Trans. Intell. V ehicles, vol. 8, no. 4, pp. 2650–2660, 2023

    work page 2023

  8. [8]

    Semantic commu nications: Principles and challenges,

    Z. Qin, X. Tao, J. Lu, W. Tong, and G. Y . Li, “Semantic commu nications: Principles and challenges,” arXiv preprint arXiv:2201.01389 , 2021

    Pith/arXiv arXiv 2021

  9. [9]

    Semantic communications for fu ture internet: Fundamentals, applications, and challenges,

    W. Y ang, H. Du, Z. Q. Liew, W. Y . B. Lim, Z. Xiong, D. Niyato, X. Chi, X. Shen, and C. Miao, “Semantic communications for fu ture internet: Fundamentals, applications, and challenges,” IEEE Commun. Surv. Tutor ., vol. 25, no. 1, pp. 213–250, Firstquarter 2023

    work page 2023

  10. [10]

    Intelligent multi-modal sen sing- communication integration: Synesthesia of machines,

    X. Cheng, H. Zhang, J. Zhang, S. Gao, S. Li, Z. Huang, L. Ba i, Z. Y ang, X. Zheng, and L. Y ang, “Intelligent multi-modal sen sing- communication integration: Synesthesia of machines,” IEEE Commun. Surv. Tutor ., vol. 26, no. 1, pp. 258–301, Firstquarter 2024

    work page 2024

  11. [11]

    A lite distributed semantic communic ation system for internet of things,

    H. Xie and Z. Qin, “A lite distributed semantic communic ation system for internet of things,” IEEE J. Sel. Areas Commun. , vol. 39, no. 1, pp. 142–153, Jan. 2021

    work page 2021

  12. [12]

    Deep learning e nabled semantic communication systems,

    H. Xie, Z. Qin, G. Y . Li, and B.-H. Juang, “Deep learning e nabled semantic communication systems,” IEEE Trans. Signal Process., vol. 69, pp. 2663–2675, Apr. 2021

    work page 2021

  13. [13]

    Performa nce limits of a deep learning-enabled text semantic communicat ion under interference,

    T. M. Getu, W. Saad, G. Kaddoum, and M. Bennis, “Performa nce limits of a deep learning-enabled text semantic communicat ion under interference,” IEEE Trans. Wireless Commun., vol. 23, no. 8, pp. 10 213– 10 228, Aug. 2024

    work page 2024

  14. [14]

    Deep jo int source- channel coding for wireless image transmission,

    E. Bourtsoulatze, D. B. Kurka, and D. G¨ und¨ uz, “Deep jo int source- channel coding for wireless image transmission,” IEEE Trans. Cogn. Commun. Netw., vol. 5, no. 3, pp. 567–579, Sept. 2019

    work page 2019

  15. [15]

    W ireless image transmission using deep source channel coding with at tention modules,

    J. Xu, B. Ai, W. Chen, A. Y ang, P . Sun, and M. Rodrigues, “W ireless image transmission using deep source channel coding with at tention modules,” IEEE Trans. Circuits Syst. Video Technol. , vol. 32, no. 4, pp. 2315–2328, Apr. 2021

    work page 2021

  16. [16]

    WIT T: A wireless image transmission transformer for semantic comm unications,

    K. Y ang, S. Wang, J. Dai, K. Tan, K. Niu, and P . Zhang, “WIT T: A wireless image transmission transformer for semantic comm unications,” in Proc. IEEE Int. Conf. Acoust. Speech Signal Process. , Rhodes Island, Greece,, Jun. 2023, pp. 1–5

    work page 2023

  17. [17]

    Task-oriented explainable semantic commun ications,

    S. Ma, W. Qiao, Y . Wu, H. Li, G. Shi, D. Gao, Y . Shi, S. Li, an d N. Al-Dhahir, “Task-oriented explainable semantic commun ications,” IEEE Trans. Wireless Commun. , vol. 22, no. 12, pp. 9248–9262, Dec. 2023

    2023

  18. [18]

    Classification-oriented semant ic wireless com- munications,

    E. Kutay and A. Y ener, “Classification-oriented semant ic wireless com- munications,” in Proc. IEEE Int. Conf. Acoustics Speech Signal Process. , Seoul, Korea (South), Apr. 2024, pp. 9096–9100

    work page 2024

  19. [19]

    Classification-driven discrete neural represen tation learning for semantic communications,

    W. Hua, L. Xiong, S. Liu, L. Chen, X. Hong, J. F. Mota, and X. Cheng, “Classification-driven discrete neural represen tation learning for semantic communications,” IEEE Internet Things J. , vol. 11, no. 9, pp. 16 061–16 073, May 2024

    work page 2024

  20. [20]

    Semantic communications with di screte-time analog transmission: A papr perspective,

    Y . Shao and D. Gunduz, “Semantic communications with di screte-time analog transmission: A papr perspective,” IEEE Wireless Commun. Lett. , vol. 12, no. 3, pp. 510–514, Mar. 2022

    work page 2022

  21. [21]

    Google/draco: a library for compressing and decompressin g 3d geo- metric meshes and point clouds,

    F. Galligan, M. Hemmer, O. Stava, F. Zhang, and J. Brettl e, “Google/draco: a library for compressing and decompressin g 3d geo- metric meshes and point clouds,” 2018

    work page 2018

  22. [22]

    An overview of ongoing point cloud compressi on stan- dardization activities: Video-based (v-pcc) and geometry -based (g-pcc),

    D. Graziosi, O. Nakagami, S. Kuma, A. Zaghetto, T. Suzuk i, and A. Tabatabai, “An overview of ongoing point cloud compressi on stan- dardization activities: Video-based (v-pcc) and geometry -based (g-pcc),” APSIPA Transactions on Signal and Information Processing , vol. 9, p. e13, 2020

    work page 2020

  23. [23]

    Geometric modeling using octree encoding ,

    D. Meagher, “Geometric modeling using octree encoding ,” Computer graphics and image processing , vol. 19, no. 2, pp. 129–147, 1982

    work page 1982

  24. [24]

    Multidimensional binary search trees u sed for associative searching,

    J. L. Bentley, “Multidimensional binary search trees u sed for associative searching,” Communications of the ACM , vol. 18, no. 9, pp. 509–517, 1975

    work page 1975

  25. [25]

    Pointnet: Deep le arning on point sets for 3d classification and segmentation,

    C. R. Qi, H. Su, K. Mo, and L. J. Guibas, “Pointnet: Deep le arning on point sets for 3d classification and segmentation,” in Proc. IEEE Conf. Comput. Vis. Pattern Recognit. , Honolulu, HI, USA, Jul. 2017, pp. 77–85

    work page 2017

  26. [26]

    Pointnet++: Deep hierarchical feature learning on point sets in a metric space,

    C. R. Qi, L. Yi, H. Su, and L. J. Guibas, “Pointnet++: Deep hierarchical feature learning on point sets in a metric space,” Red Hook, N Y , USA, 2017, pp. 5105–5114

    work page 2017

  27. [27]

    Poin t transformer,

    H. Zhao, L. Jiang, J. Jia, P . H. Torr, and V . Koltun, “Poin t transformer,” in Proc. IEEE Int. Conf. Comput. Vis., Montreal, QC, Canada, Oct. 2021, pp. 16 239–16 248

    work page 2021

  28. [28]

    3d point cloud geometry compressio n on deep learning,

    T. Huang and Y . Liu, “3d point cloud geometry compressio n on deep learning,” in Proc. ACM Int. Conf. Multimedia , New Y ork, NY , USA, Oct. 2019, pp. 890–898

    work page 2019

  29. [29]

    Transformer and ups ampling- based point cloud compression,

    J. Zhang, G. Liu, D. Ding, and Z. Ma, “Transformer and ups ampling- based point cloud compression,” in Proc. ACM Adv. Point Cloud Comp. Process. Analysis, New Y ork, NY , USA, Oct. 2022, pp. 33–39

    work page 2022

  30. [30]

    Deep compression for dense point cloud maps,

    L. Wiesmann, A. Milioto, X. Chen, C. Stachniss, and J. Be hley, “Deep compression for dense point cloud maps,” IEEE Rob. Autom. Lett., vol. 6, no. 2, pp. 2060–2067, Apr. 2021

    work page 2060

  31. [31]

    Octattention: O ctree-based large-scale contexts model for point cloud compression,

    C. Fu, G. Li, R. Song, W. Gao, and S. Liu, “Octattention: O ctree-based large-scale contexts model for point cloud compression,” i n Proc. AAAI Conf. Artif. Intell. , vol. 36, no. 1, Jun. 2022, pp. 625–633

    work page 2022

  32. [32]

    Densit y- preserving deep point cloud compression,

    Y . He, X. Ren, D. Tang, Y . Zhang, X. Xue, and Y . Fu, “Densit y- preserving deep point cloud compression,” in Proc. IEEE Conf. Comput. Vis. Pattern Recognit. , New Orleans, LA, USA, Jun. 2022, pp. 2323– 2332

    work page 2022

  33. [33]

    C. E. Shannon and W. Weaver, The Mathematical Theory of Communi- cation. The University of Illinois Press, 1949

    work page 1949

  34. [34]

    Deep joint source-cha nnel coding for csi feedback: An end-to-end approach,

    J. Xu, B. Ai, N. Wang, and W. Chen, “Deep joint source-cha nnel coding for csi feedback: An end-to-end approach,” IEEE J. Sel. Areas Commun. , vol. 41, no. 1, pp. 260–273, Jan. 2023

    work page 2023

  35. [35]

    Task-oriented multi-user semantic communications,

    H. Xie, Z. Qin, X. Tao, and K. B. Letaief, “Task-oriented multi-user semantic communications,” IEEE J. Sel. Areas Commun. , vol. 40, no. 9, pp. 2584–2597, Sept. 2022

    work page 2022

  36. [36]

    A unified multi- task semantic communication system with domain adaptation,

    G. Zhang, Q. Hu, Z. Qin, Y . Cai, and G. Y u, “A unified multi- task semantic communication system with domain adaptation,” in Proc. IEEE Global Comm. Conf. , Rio de Janeiro, Brazil, Dec. 2022, pp. 3971–3976

    work page 2022

  37. [37]

    Multi-modal fus ion- based multi-task semantic communication system,

    Z. Zhu, R. Zhang, X. Cheng, and L. Y ang, “Multi-modal fus ion- based multi-task semantic communication system,” arXiv preprint arXiv:2407.00964, 2024

    arXiv 2024

  38. [38]

    Wireless 3d point cloud delivery using deep graph neural networks,

    T. Fujihashi, T. Koike-Akino, S. Chen, and T. Watanabe, “Wireless 3d point cloud delivery using deep graph neural networks,” in IEEE Int. Conf. Commun. , Montreal, QC, Canada, Jun. 2021, pp. 1–6

    work page 2021

  39. [39]

    Semantic com mu- nications system with model division multiple access and co ntrollable coding rate for point cloud,

    X. Liu, H. Liang, Z. Bao, C. Dong, and X. Xu, “Semantic com mu- nications system with model division multiple access and co ntrollable coding rate for point cloud,” arXiv preprint arXiv:2307.06027 , 2023

    Pith/arXiv arXiv 2023

  40. [40]

    Over-the-Air Learning-based Geometry Point Cloud Transmission

    C. Bian, Y . Shao, and D. Gunduz, “Wireless point cloud tr ansmission,” arXiv preprint arXiv:2306.08730 , 2023

    work page internal anchor Pith review Pith/arXiv arXiv 2023

  41. [41]

    Shapenet: An information- rich 3d model repository,

    A. X. Chang, T. Funkhouser, L. Guibas, P . Hanrahan, Q. Hu ang, Z. Li, S. Savarese, M. Savva, S. Song, H. Su et al., “Shapenet: An information- rich 3d model repository,” arXiv preprint arXiv:1512.03012 , 2015

    Pith/arXiv arXiv 2015

  42. [42]

    3d shapenets: A deep representation for volumetric shapes,

    Z. Wu, S. Song, A. Khosla, F. Y u, L. Zhang, X. Tang, and J. X iao, “3d shapenets: A deep representation for volumetric shapes,” i n Proc. IEEE Conf. Comput. Vis. Pattern Recognit. , Boston, MA, USA, Jun. 2015, pp. 1912–1920

    work page 2015

  43. [43]

    Estimating o r propagating gradients through stochastic neurons for conditional comp utation,

    Y . Bengio, N. L´ eonard, and A. Courville, “Estimating o r propagating gradients through stochastic neurons for conditional comp utation,” arXiv preprint arXiv:1308.3432, 2013

    Pith/arXiv arXiv 2013

  44. [44]

    Differentiable soft quantization: Bridging full-precis ion and low-bit neural networks,

    R. Gong, X. Liu, S. Jiang, T. Li, P . Hu, J. Lin, F. Y u, and J. Y an, “Differentiable soft quantization: Bridging full-precis ion and low-bit neural networks,” in Proc. IEEE Conf. Comput. Vis. Pattern Recognit. , Seoul, Korea (South), Oct.-Nov. 2019, pp. 4852–4861

    work page 2019

  45. [45]

    Unde rstanding straight-through estimator in training activation quanti zed neural nets,

    P . Yin, J. Lyu, S. Zhang, S. Osher, Y . Qi, and J. Xin, “Unde rstanding straight-through estimator in training activation quanti zed neural nets,” arXiv preprint arXiv:1903.05662 , 2019

    Pith/arXiv arXiv 1903

  46. [46]

    Opv2v: An open benchmark dataset and fusion pipeline for perception w ith vehicle- 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 w ith vehicle- to-vehicle communication,” in Proc. IEEE Int. Conf. Robot. Autom. , Philadelphia, PA, USA, May 2022, pp. 2583–2589

    work page 2022

This paper was first reviewed by deepseek-v4-flash on August 4, 2026.