Pith. sign in

REVIEW 4 major objections 6 minor 31 references

Channel-adaptive Cross-modal Generative Semantic Communication for Point Cloud Transmission

T0 review · 4 major / 6 minor · reviewed 2026-08-07 · deepseek-v4-flash

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

Pith's one-line read The paper claims that point clouds can be transmitted as a few hundred analog semantic symbols—over 1000x fewer than octree coding—by using 2D images as local keypoint guides and a diffusion decoder to regenerate the 3D shape.

desk verdict Solid integration paper with a real ablation; the headline -20 dB/1000x numbers rest on a single object, so treat them as a proof of concept pending multi-object evaluation. read the letter →

arxiv 2506.03211 v1 pith:B6ONWFR2 submitted 2025-06-03 cs.CV cs.NIeess.SP

Wanting Yang , Zehui Xiong , Qianqian Yang , Ping Zhang , Merouane Debbah , Rahim Tafazolli This is my paper
classification cs.CVcs.NIeess.SP
keywords semanticcommunicationpointcloudtransmissiongenerativediffusionmodelcross-modalextractionjointsource-channelcodingkeypointdetectionchanneladaptationanalog
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 GenSeC-PC, a semantic communication system that transmits 3D point clouds over wireless channels by sending compact analog semantic features instead of compressed raw geometry. The central claim is that cross-modal semantic extraction—using multi-view 2D images of the object, which are never transmitted, to locate keypoints in the point cloud—produces a richer prompt for a generative diffusion decoder, improving reconstruction over the point-cloud-only PointDif baseline on all four reported metrics (CD, HD, EMD, MSE). Because the semantic representation travels fully in analog form and the decoder tolerates distortion, the system reportedly beats conventional octree-based digital transmission by over three orders of magnitude in transmitted symbols while preserving recognizable object structure even at SNR near -20 dB. The paper also claims a channel-adaptive joint source-channel coding scheme in which only the encoder needs SNR and bandwidth feedback, plus millisecond-level decoding via DDIM and Rectified Diffusion. If correct, this points a practical route around the cliff effect and the lossless-side-information overhead that has limited earlier point cloud semantic communication.

What carries the argument

Three components carry the argument. (1) Cross-modal semantic extraction: multi-view 2D images are processed by a pretrained DINOv2 vision encoder, the per-view features are back-projected into 3D and matched against pre-stored few-shot keypoint features following the B2-3D method to detect keypoints in the point cloud; keypoint-aware farthest point sampling (KP-FPS) then builds patches with $G \times S = 2 \times N$ oversampling, random masking forces global reasoning, and keypoint embeddings are injected into the tokens of keypoint-centered patches before aggregation into a single semantic feature. (2) Channel-adaptive analog JSCC: an asymmetric encoder/decoder pair, pruned from a larger base model, in which SNR Adaptation Components (SACs) and Rate Adaptation Components (RACs) reweight the feature through gating functions conditioned on average SNR and available bandwidth, and a Conditional Multi-Branch Component (CMBC) selects the rate without explicit rate signaling; on the receive side, Feature Adaptation Components (FACs) compensate for feature-dependent semantic loss. (3) A generative diffusion decoder built from controlled-PC (C-PC) blocks that condition each denoising step on the recovered semantic feature and the time step, accelerated by DDIM deterministic sampling and retrained with Rectified Diffusion on matched noise-sample pairs to reach millisecond-level inference.

What would settle it

Run GenSeC-PC on object classes with no distinctive keypoints (smooth, symmetric shapes) or with the 2D-image pathway disabled while keeping rate and SNR fixed: if CD and HD stay essentially unchanged relative to the PointDif baseline, the cross-modal extraction is not doing the claimed work. Conversely, transmit one object by octree coding with an optimized entropy coder and channel code consuming the same total symbol budget as GenSeC-PC at rate 32: if octree matches or beats its CD at that budget, the over-1000x spectral-efficiency claim collapses.

Watch

Extended reading notes

Core claim

On the paper's own terms, the central discovery is that point cloud transmission is better framed as a conditional generation problem than a compression problem: a semantic encoder extracts a compact, keypoint-guided feature from a partial and possibly noisy point cloud, fusing it with features back-projected from multi-view 2D images, and a diffusion-based decoder built on PointDif regenerates the complete cloud from that feature as a prompt. The transmitted representation is a fully analog semantic symbol vector whose length is adapted to the available bandwidth and whose entries are reweighted by the average SNR at the encoder only; the decoder is a fixed, feedback-free mirror that recovers the feature and feeds it to the diffusion backbone. GenSeC-PC thereby removes the need for lossless digital transmission of side information that prior JSCC-based point cloud systems required, and the authors report that a single object can be sent with over 1000x fewer symbols than octree coding while remaining structurally recognizable at SNR as low as -20 dB at rate 896. The cross-modal keypoint injection (IaKP-PointDif) and the dual-metric fine-tuning (MSE plus CD) are each shown by ablation to improve reconstruction across object categories.

Load-bearing premise

The whole advantage rests on the transmitter having both the 3D point cloud and multi-view 2D images of the same object, plus a pre-stored library of few-shot keypoint features for that object class and an object detector to call it up; without these, the semantic encoder cannot produce the keypoint-guided features the diffusion decoder was trained to condition on, and the bandwidth advantage over octree transmission disappears.

Editorial extensions

If this is right

  • Over 1000x fewer transmitted symbols than octree-based coding at comparable reconstruction quality would make point cloud delivery practical in spectrum-limited settings such as autonomous driving and extended reality offloading.
  • Because the decoder tolerates noisy and incomplete source clouds, the transmitter no longer needs a clean, complete scan; a partial LiDAR view with occlusions can still yield a full reconstruction.
  • The asymmetric feedback design—only the encoder conditions on SNR and bandwidth—removes the need to send channel state to the receiver, simplifying the protocol stack.
  • Millisecond-level decoding via DDIM plus Rectified Diffusion makes diffusion-based reconstruction compatible with real-time interactive applications rather than offline rendering.
  • Graceful degradation tied to rate, with object structure preserved even near -20 dB SNR at rate 896, offers a predictable quality-versus-bandwidth operating curve instead of the cliff effect of digital decoding.

Reading between the lines

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

  • A natural stress test the paper does not run: disable the 2D-image pathway entirely and transmit with the same rate and SNR. The paper's own ablation implies the gains concentrate in the keypoint-guided cross-modal pathway, so on smooth, featureless shapes where keypoints are not discriminative the system should degrade toward its PointDif baseline.
  • The diffusion decoder accepts prompts of varying quality, so a progressive or rateless reading follows implicitly: a receiver could reconstruct a coarse shape from the first few received symbols and refine as more arrive, an operation mode the paper does not explore.
  • The >1000x spectral-efficiency figure counts wireless symbols only; a fair end-to-end accounting would also price the transmitter-side costs of DINOv2 computation and the per-class keypoint feature dictionary that the authors list as limitations.
  • A deployment-shaped consequence: a fleet of sensors could share a cached keypoint-prior library for known object classes, but the advantage shrinks to that of a plain point-cloud encoder when the class is novel or the cache is stale.
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

4 major / 6 minor

Summary. The paper proposes GenSeC-PC, a point-cloud semantic communication system that combines a cross-modal semantic encoder (point cloud plus multi-view 2D images as non-transmitted side information), an SNR- and rate-adaptive asymmetric JSCC encoder/decoder, and a diffusion-based generative decoder. The semantic encoder uses B2-3D keypoint detection and random patch masking; the decoder is based on the PointDif backbone with DDIM/RD acceleration. Training uses pre-training with MSE, fine-tuning with MSE+CD, and JSCC training with MSE on transmitted features plus CD on reconstructed clouds. Experiments on KeypointNet and ShapeNet compare against PointDif and octree-based conventional transmission, reporting lower CD/HD/EMD, high spectral efficiency, and robustness to low SNR, keypoint errors, and unseen shapes.

Significance. If the central claims hold, the paper contributes a practical integration of cross-modal 3D semantics with channel-adaptive JSCC and generative decoding, with an interesting asymmetric design that keeps SNR/rate feedback on the encoder side only. The internal ablation (IaKP-PointDif vs PointDif, Fig. 7) is genuine and shows consistent gains across categories and metrics, supporting the cross-modal benefit. The channel simulations show the expected monotonic degradation with SNR/rate, and the use of a full holdout evaluation on unseen ShapeNet shapes is a positive feature. The code and trained models are promised publicly. However, several load-bearing evaluation gaps—single-object transmission results, missing test-time evaluation for incomplete/noisy source clouds, and qualitative-only unseen-shape evidence—currently limit the strength of the headline claims.

major comments (4)
  1. [Section IV-B-2 (text preceding Fig. 9)] All quantitative transmission results, including the CD/HD/EMD surfaces in Fig. 9, the examples in Fig. 10, and the octree comparison in Fig. 12, are based on a single object transmitted 100 times. Since Section IV-B-1 and Fig. 7 show substantial reconstruction-quality variation across categories, the abstract's headline claims that the system preserves object structure at SNR = -20 dB and achieves over 1000x fewer symbols than octree transmission are not established for the general case. Please report transmission metrics averaged over a sample of objects spanning multiple categories, with per-object distributions, and state explicitly whether the -20 dB and 1000x figures hold across that sample.
  2. [Abstract and Section III-A/III-D; Section IV-B-3] The paper repeatedly claims robustness to noisy or incomplete source point clouds (Abstract, Section I, Section III-A). The only mechanism supporting this claim is random patch masking applied during training, with a visible-patch fraction of 80% (V = floor(G*iota) with iota = 0.8). No test-time experiment feeds incomplete or noisy source point clouds to the trained pipeline. Section IV-B-3 evaluates robustness to inaccurate keypoint detection and to unseen ShapeNet shapes, but not to incomplete/noisy source clouds. To support the stated claim, add a direct evaluation where the source cloud is perturbed (e.g., random point removal, sensor noise, or masked regions) at inference, and report CD/HD/EMD across rates and SNRs.
  3. [Section IV-B-3 (Fig. 14)] The robustness claim for previously unseen objects is supported only by qualitative visualizations. No CD/HD/EMD numbers are reported for the typical and uncommon ShapeNet shapes in Fig. 14, nor is a comparison with PointDif or another baseline given. Since 'previously unseen objects' is a headline contribution, please provide quantitative metrics on a held-out set of ShapeNet objects and compare against a non-cross-modal baseline.
  4. [Section III-B and Section IV-A] The channel-adaptive JSCC is described as 'derived via pruning from a larger base model' and trained via a 'pretraining-pruning pipeline,' but the base architecture, pruning criterion, and the resulting pruned structure are not specified. Only the final module counts and rate/SNR level sets are given. Because the asymmetric JSCC design is a central contribution, this omission prevents reproduction and makes the 'streamlined' claim difficult to verify. Please provide the base model description, the pruning procedure, and a comparison of complexity before and after pruning.
minor comments (6)
  1. [Equation (7)] Equation (7) appears to contain a typo: the position embeddings {E^v_i} are written as a function of the masked patches {P^m_i}, although the surrounding text states they are derived from the centers of visible patches. Please correct the equation or the text to make the two consistent.
  2. [Section III-A and Section IV-A] The quantity iota is called the masking ratio, but the definition V = floor(G*i) with iota = 0.8 yields about 51 visible patches out of 64, i.e., iota is the visible fraction, not the masked fraction. Please rename or clarify to avoid confusion.
  3. [Figure 8] Figure 8 labels a comparison as 'Baseline [34]', but the reference list ends at [31]. This looks like a dangling citation; please add the reference or replace it with the intended method name.
  4. [Section IV-B-2 (discussion after Fig. 10)] The text mentions 'the critical SNR points for rates of 640 and 265'; the second rate is presumably 256. Please correct this typo.
  5. [Section IV-B-3] The text says transmission takes place 'at a rate of 512 points per cloud' (e.g., in the sentence before Fig. 14); this should be '512 symbols per cloud' to match the rate definition used elsewhere.
  6. [Abstract and Section I] The phrase 'rectified denoising diffusion implicit models' is imprecise: the paper actually combines DDIM sampling with a separately trained rectified diffusion model. Please rephrase to distinguish the two techniques clearly.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity; GenSeC-PC is an end-to-end learned pipeline with genuine holdout evaluation and independent external priors.

full rationale

GenSeC-PC is an end-to-end learned system, not a closed-form derivation whose outputs are fixed by fitted inputs. The cross-modal keypoint features come from the external B2-3D keypoint detector [22] and the decoder is built on the external PointDif diffusion backbone [23]; neither is a self-citation of the present authors, and the claimed novelty lies in the SNR/rate-adaptive JSCC architecture, the cross-modal conditioning, and the analog-transmission design. The only train/eval overlap is that CD appears both in the dual-metric fine-tuning loss (Eq. 28-29) and in the evaluation metrics (Eq. 2); this is a standard alignment of training objective and evaluation metric, and the paper's relative comparisons are informative because the PointDif baseline is fine-tuned under the identical dual-metric protocol. The unseen-shape experiments on ShapeNet (Sec. IV-B-3, Fig. 14) are genuine holdouts, and the transmission curves in Figs. 9 and 12, while restricted to a single object repeated 100 times, are empirical measurements rather than quantities forced by a fitted parameter or by a self-citation chain. No equation in the paper reduces to its own input, no prediction is statistically forced by an earlier fit, and the paper's stated limitations (e.g., reliance on B2-3D keypoints and object detection capability) are openly acknowledged rather than hidden. Therefore, the derivation chain is self-contained for the purposes of a circularity analysis.

Assumptions & free parameters 3 free parameters · 4 assumptions · 0 invented entities

The central claim rests on the availability of multi-view image side information and a per-class keypoint prior at the transmitter, which are domain assumptions not guaranteed in typical point cloud transmission scenarios. The network hyperparameters are chosen by hand, and no sensitivity study is provided.

free parameters (3)
  • masking ratio = 0.8
    Set by hand in Section IV.A; controls the amount of input corruption simulated during training and affects the robustness claim.
  • feature dimensions = d1=384, d2=512, d=1024
    Chosen capacity; no sensitivity analysis.
  • rate set = [1024, 896, 768, 640, 512, 384, 256, 128, 64, 32]
    Discrete available bandwidth levels; the receiver infers rate from symbol length.
assumptions (4)
  • domain assumption The source has both a 3D point cloud and multi-view 2D images of the same object.
    Section II.A states the system relies on two sensors, LiDAR and a drone camera. If no camera views exist, keypoint detection cannot be performed.
  • ad hoc to paper The source has object detection capability and a pre-stored feature repository for the object's class (few-shot keypoint features).
    Section III.A says B2-3D requires this, and the Conclusion lists it as a limitation.
  • domain assumption The diffusion decoder pre-trained on KeypointNet generalizes to the target point cloud distribution.
    Standard ML transfer assumption; tested only on ShapeNet models of the same categories.
  • domain assumption Channel state information (average SNR and available bandwidth) is available to the sender, and the receiver knows the transmitted symbol length without extra signaling.
    Used in Sections III.B; the length inference is an implicit signaling mechanism.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Channel-adaptive Cross-modal Generative Semantic Communication for Point Cloud Transmission." pith.science (2026). https://pith.science/paper/B6ONWFR2

@misc{pith2026250603211,
  author       = {Pith},
  title        = {Pith review of: Channel-adaptive Cross-modal Generative Semantic Communication for Point Cloud Transmission},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/B6ONWFR2}},
  note         = {Machine review of arXiv:2506.03211}
}
read the original abstract

With the rapid development of autonomous driving and extended reality, efficient transmission of point clouds (PCs) has become increasingly important. In this context, we propose a novel channel-adaptive cross-modal generative semantic communication (SemCom) for PC transmission, called GenSeC-PC. GenSeC-PC employs a semantic encoder that fuses images and point clouds, where images serve as non-transmitted side information. Meanwhile, the decoder is built upon the backbone of PointDif. Such a cross-modal design not only ensures high compression efficiency but also delivers superior reconstruction performance compared to PointDif. Moreover, to ensure robust transmission and reduce system complexity, we design a streamlined and asymmetric channel-adaptive joint semantic-channel coding architecture, where only the encoder needs the feedback of average signal-to-noise ratio (SNR) and available bandwidth. In addition, rectified denoising diffusion implicit models is employed to accelerate the decoding process to the millisecond level, enabling real-time PC communication. Unlike existing methods, GenSeC-PC leverages generative priors to ensure reliable reconstruction even from noisy or incomplete source PCs. More importantly, it supports fully analog transmission, improving compression efficiency by eliminating the need for error-free side information transmission common in prior SemCom approaches. Simulation results confirm the effectiveness of cross-modal semantic extraction and dual-metric guided fine-tuning, highlighting the framework's robustness across diverse conditions, including low SNR, bandwidth limitations, varying numbers of 2D images, and previously unseen objects.

Figures

Figures reproduced from arXiv: 2506.03211 by the authors.

Figure 1
Figure 1. Illustration of channel-adaptive cross-modal GenSeC-PC framework. [PITH_FULL_IMAGE:figures/full_fig_p003_1.png] view at source ↗
Figure 2
Figure 2. Illustration of cross-modal semantic extraction. (a) [PITH_FULL_IMAGE:figures/full_fig_p004_2.png] view at source ↗
Figure 3
Figure 3. Asymmetric architecture of channel-adaptive JSCC encoder and decoder. [PITH_FULL_IMAGE:figures/full_fig_p005_3.png] view at source ↗
Figures from the paper (11 more)
Figure 5
Figure 5. Figure 5: Backbone of the PC diffusion. Then, the obtained Fˆ (IFea) s is fed into IFea FACs in order. FAC is similar to SAC and RAC, except that the SNR or rate feature is replaced by the input feature F (i+1) s of FAC i. It can be highlighted as Wi |1×d| = 2 × GateS D  Fˆ(i+1…
Figure 6
Figure 6. Figure 6: Training flow of GenSeC-PC. by ϵθ, consists of multiple controlled-PC (C-PC) blocks [27], following a Decoder-Encoder architecture that structurally re￾sembles an “inverse U-Net”. For each C-PC block, the feature from previous block Hl−1 is first fed into an MLP, perfo…
Figure 7
Figure 7. Figure 7: Performance comparison between IaKP-PointDif and PointDif. [PITH_FULL_IMAGE:figures/full_fig_p008_7.png]
Figure 8
Figure 8. Figure 8: Comparison of visualization results between IaKP-PointDif and PointDif. [PITH_FULL_IMAGE:figures/full_fig_p009_8.png]
Figure 9
Figure 9. Figure 9: Communication performance of GenSeC-PC under AWGN and Rayleigh channels with varying SNR and rate. [PITH_FULL_IMAGE:figures/full_fig_p010_9.png]
Figure 10
Figure 10. Figure 10: Typical visualization of transmission results under [PITH_FULL_IMAGE:figures/full_fig_p010_10.png]
Figure 12
Figure 12. Figure 12: Performance comparison between GenSeC-PC and traditional method under AWGN channel, averaged over 100 [PITH_FULL_IMAGE:figures/full_fig_p011_12.png]
Figure 13
Figure 13. Figure 13: Decoding acceleration with DDIM and RD. fails entirely under low-SNR conditions. Although reducing the depth to 4 enables successful decoding, the resulting distortion remains extremely high. In contrast, the proposed GenSeC-PC achieves significantly lower CD across v…
Figure 14
Figure 14. Figure 14: Performance evaluation on the unseen models in [PITH_FULL_IMAGE:figures/full_fig_p011_14.png]
Figure 15
Figure 15. Figure 15: Visualization results of keypoint detection, using different numbers of view images for acquisition and PC reconstruction [PITH_FULL_IMAGE:figures/full_fig_p012_15.png]
Figure 16
Figure 16. Figure 16: Average performance of keypoint detection, using [PITH_FULL_IMAGE:figures/full_fig_p012_16.png]

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

31 extracted references · 23 canonical work pages

  1. [1]

    A review of algorithms for filtering the 3d point cloud,

    X.-F. Han, J. S. Jin, M.-J. Wang, W. Jiang, L. Gao, and L. Xiao, “A review of algorithms for filtering the 3d point cloud,” Signal Processing: Image Communication, vol. 57, pp. 103–112, 2017

    work page 2017

  2. [2]

    Pstnet: Point spatio-temporal convolution on point cloud sequences,

    H. Fan, X. Yu, Y . Ding, Y . Yang, and M. Kankanhalli, “Pstnet: Point spatio-temporal convolution on point cloud sequences,” in International Conference on Learning Representations , 2021

    work page 2021

  3. [3]

    3d point cloud compression: A survey,

    C. Cao, M. Preda, and T. Zaharia, “3d point cloud compression: A survey,” in Proceedings of the 24th International Conference on 3D Web Technology, 2019, pp. 1–9

    work page 2019

  4. [4]

    Octree-based point-cloud compression

    R. Schnabel and R. Klein, “Octree-based point-cloud compression.” PBG@ SIGGRAPH, vol. 3, no. 3, 2006

    work page 2006

  5. [5]

    A semantic com- munication system for point cloud,

    X. Liu, H. Liang, Z. Bao, C. Dong, and X. Xu, “A semantic com- munication system for point cloud,” IEEE Transactions on Vehicular Technology, vol. 74, no. 1, pp. 894–910, 2025

    work page 2025

  6. [6]

    Adaptive graph convolution for point cloud analysis,

    H. Zhou, Y . Feng, M. Fang, M. Wei, J. Qin, and T. Lu, “Adaptive graph convolution for point cloud analysis,” in Proceedings of the IEEE/CVF international conference on computer vision , 2021, pp. 4965–4974

    work page 2021

  7. [7]

    Point-bert: Pre-training 3d point cloud transformers with masked point modeling,

    X. Yu, L. Tang, Y . Rao, T. Huang, J. Zhou, and J. Lu, “Point-bert: Pre-training 3d point cloud transformers with masked point modeling,” in Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, 2022, pp. 19 313–19 322

    2022

  8. [8]

    Semantic-aware transmission for robust point cloud classification,

    T. Han, K. Chi, Q. Yang, and Z. Shi, “Semantic-aware transmission for robust point cloud classification,” in GLOBECOM 2023 - 2023 IEEE Global Communications Conference , 2023, pp. 7617–7622

    work page 2023

Show all 31 references
  1. [9]

    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,” in Proceedings of the 31st International Conference on Neural Information Processing Systems, ser. NIPS’17. Red Hook, NY , USA: Curran Associates Inc., 2...

    2017

  2. [10]

    Semantic communication for efficient point cloud transmission,

    S. Xie, Q. Yang, Y . Sun, T. Han, Z. Yang, and Z. Shi, “Semantic communication for efficient point cloud transmission,” in GLOBECOM 2024 - 2024 IEEE Global Communications Conference, 2024, pp. 2948– 2953

    2024

  3. [11]

    Inception-v4, inception-resnet and the impact of residual connections on learning,

    C. Szegedy, S. Ioffe, V . Vanhoucke, and A. A. Alemi, “Inception-v4, inception-resnet and the impact of residual connections on learning,” in Proceedings of the Thirty-First AAAI Conference on Artificial Intelli- gence, ser. AAAI’17. AAAI Press, 2017, p. 4278–4284

    2017

  4. [12]

    Deep joint source-channel coding for wireless point cloud transmission,

    C. Zhang, M. Liu, W. Huang, Y . Xu, Y . Xu, and D. He, “Deep joint source-channel coding for wireless point cloud transmission,” in ICASSP 2025 - 2025 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP) , 2025, pp. 1–5

    2025

  5. [13]

    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 ICC 2021 - IEEE International Conference on Communications , 2021, pp. 1–6

    2021

  6. [14]

    Rateless deep joint source channel coding for 3d point cloud,

    S. Ibuki, T. Okamoto, T. Fujihashi, T. Koike-Akino, and T. Watanabe, “Rateless deep joint source channel coding for 3d point cloud,” IEEE Access, vol. 13, pp. 39 585–39 599, 2025

    2025

  7. [15]

    Shapenet: An information- rich 3d model repository,

    A. X. Chang, T. Funkhouser, L. Guibas, P. Hanrahan, Q. Huang, 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

    2015 arXiv

  8. [16]

    Spatial- temporal semantic communications for point cloud-based volumetric media,

    J. Men, N. Wang, Y . Ma, P. Qian, C. Udora, and M. Nilsson, “Spatial- temporal semantic communications for point cloud-based volumetric media,” in ICC 2024 - IEEE International Conference on Communi- cations, 2024, pp. 4704–4710

    2024

  9. [17]

    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. Yang, and P. Zhang, “Iscom: Interest-aware semantic communication scheme for point cloud video streaming on metaverse xr devices,” IEEE Journal on Selected Areas in Communications , vol. 42, no. 4, pp. 1003–1021, 2023

    2023

  10. [18]

    A semantic communication system for real-time 3d reconstruction tasks,

    J. Zhang, L. Guo, K. Zhu, and H. Qiu, “A semantic communication system for real-time 3d reconstruction tasks,” in 2024 8th Interna- tional Conference on Communication and Information Systems (ICCIS) . IEEE, 2024, pp. 150–155

    2024

  11. [19]

    Large generative model assisted 3d semantic communication,

    F. Jiang, Y . Peng, L. Dong, K. Wang, K. Yang, C. Pan, and X. You, “Large generative model assisted 3d semantic communication,” arXiv preprint arXiv:2403.05783, 2024

    2024 arXiv

  12. [20]

    Multi-modal semantic communication with point cloud diffusion,

    W. Feng and Z. Shi, “Multi-modal semantic communication with point cloud diffusion,” in ICASSP 2025 - 2025 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP) , 2025, pp. 1–5

    2025

  13. [21]

    Pc2: Projection- conditioned point cloud diffusion for single-image 3d reconstruction,

    L. Melas-Kyriazi, C. Rupprecht, and A. Vedaldi, “Pc2: Projection- conditioned point cloud diffusion for single-image 3d reconstruction,” in Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2023, pp. 12 923–12 932

    2023

  14. [22]

    Back to 3d: Few-shot 3d keypoint detection with back-projected 2d features,

    T. Wimmer, P. Wonka, and M. Ovsjanikov, “Back to 3d: Few-shot 3d keypoint detection with back-projected 2d features,” in Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2024

    2024

  15. [23]

    Point cloud pre-training with diffusion models,

    X. Zheng, X. Huang, G. Mei, Y . Hou, Z. Lyu, B. Dai, W. Ouyang, and Y . Gong, “Point cloud pre-training with diffusion models,” in Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2024, pp. 22 935–22 945

    2024

  16. [24]

    DINOv2: Learning robust visual features without supervision,

    M. Oquab, T. Darcet, T. Moutakanni, H. V . V o, M. Szafraniec, V . Khalidov, P. Fernandez, D. HAZIZA, F. Massa, A. El-Nouby, M. Assran, N. Ballas, W. Galuba, R. Howes, P.-Y . Huang, S.-W. Li, I. Misra, M. Rabbat, V . Sharma, G. Synnaeve, H. Xu, H. Jegou, J. Mairal, P. Labatut,...

    2024

  17. [25]

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

    C. R. Qi, H. Su, K. Mo, and L. J. Guibas, “Pointnet: Deep learning on point sets for 3d classification and segmentation,” in Proceedings of the IEEE conference on computer vision and pattern recognition , 2017, pp. 652–660

    2017

  18. [26]

    Digital-sc: Digital semantic communication with adaptive network split and learned non-linear quantization,

    L. Guo, W. Chen, Y . Sun, and B. Ai, “Digital-sc: Digital semantic communication with adaptive network split and learned non-linear quantization,” IEEE Transactions on Cognitive Communications and Networking, pp. 1–1, 2024

    2024

  19. [27]

    Diffusion probabilistic models for 3d point cloud generation,

    S. Luo and W. Hu, “Diffusion probabilistic models for 3d point cloud generation,” in Proceedings of the IEEE/CVF conference on computer vision and pattern recognition , 2021, pp. 2837–2845

    2021

  20. [28]

    Denoising diffusion probabilistic models,

    J. Ho, A. Jain, and P. Abbeel, “Denoising diffusion probabilistic models,” Advances in neural information processing systems , vol. 33, pp. 6840– 6851, 2020

    2020

  21. [29]

    Denoising diffusion implicit models,

    J. Song, C. Meng, and S. Ermon, “Denoising diffusion implicit models,” in International Conference on Learning Representations , 2021. [Online]. Available: https://openreview.net/forum?id=St1giarCHLP

    2021

  22. [30]

    Rectified diffusion: Straightness is not your need in rectified flow,

    F.-Y . Wang, L. Yang, Z. Huang, M. Wang, and H. Li, “Rectified diffusion: Straightness is not your need in rectified flow,” in The Thirteenth International Conference on Learning Representations, 2025. [Online]. Available: https://openreview.net/forum?id=nEDToD1R8M

    2025

  23. [31]

    Keypointnet: A large-scale 3d keypoint dataset aggregated from numer- ous human annotations,

    Y . You, Y . Lou, C. Li, Z. Cheng, L. Li, L. Ma, C. Lu, and W. Wang, “Keypointnet: A large-scale 3d keypoint dataset aggregated from numer- ous human annotations,” in 2020 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) , 2020, pp. 13 644–13 653

    2020

Pith tools

Reviewed August 7, 2026 · model on record in the stance chip above.