Pith. sign in

REVIEW 6 major objections 5 minor 1 cited by

Adaptive 3D Gaussian Splatting Video Streaming: Visual Saliency-Aware Tiling and Meta-Learning-Based Bitrate Adaptation

T0 review · 6 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 The paper claims that saliency-aware adaptive tiling, motion-categorized dynamic encoding, and meta-reinforcement-learning bitrate control together make 3DGS video streamable over 4G and 5G networks, maximizing quality of experience.

desk verdict Plausible system integration for 3DGS streaming, but the evaluation optimizes an unvalidated PSNR/SSIM proxy and skips the most relevant baseline. read the letter →

arxiv 2507.14454 v1 pith:NVXVAPKN submitted 2025-07-19 cs.CV cs.MMeess.IV

Han Gong , Qiyue Li , Jie Li , Zhi Liu This is my paper
classification cs.CVcs.MMeess.IV
keywords 3DGaussiansplattingvolumetricvideostreamingadaptivebitratevisualsaliencytilingmeta-reinforcementlearningqualityofexperiencedeformationfields
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 seeks to make 3D Gaussian splatting (3DGS) video—an emerging volumetric video format that captures scenes as clouds of 3D Gaussian ellipsoids—streamable over today's 4G and 5G networks. It claims that a complete pipeline of three components achieves that: saliency-aware adaptive tiling that shapes transmission around visual attention, tile-based dynamic encoding that compresses motion with deformation fields and prunes Gaussians into multiple quality levels, and a meta-reinforcement-learning bitrate controller trained to maximize a 3DGS-specific quality-of-experience (QoE) metric. The authors report that this pipeline outperforms state-of-the-art tiling and adaptive-bitrate methods across four network scenarios in simulations. If the claim holds, immersive 3D video becomes practical to deliver to headsets and phones without the prohibitive per-frame data volumes that currently block 3DGS streaming.

What carries the argument

Two objects carry the argument. The first is the rendering-weight importance metric $w_i = \sigma_i \sqrt{\det(\Sigma_i)}$, the opacity of a Gaussian primitive times the square root of the determinant of its covariance matrix. It defines the sampling probability $p_{\text{sample}}(i) = w_i / \sum_n w_n$ used both to feed the saliency network and to decide which Gaussians are pruned when a tile is compressed to a lower quality tier, so the same quantity aligns attention estimation with compression. The second is the layered QoE function of Eqs. (38)-(42), whose geometric term is PSNR weighted by a salience and visibility factor $\Phi_{t,k}$ and whose rendering term is SSIM weighted by an occlusion factor $\Psi_{t,k}$; the meta-RL agent's reward further adds stall time, stall frequency, and a temporal smoothness penalty. Around these two objects the system builds a spatiotemporal saliency detector with a Local Discrepancy Catcher, a motion-classified deformation-field encoder, and a meta-reinforcement-learning policy with task embeddings and feature-wise linear modulation.

What would settle it

A subjective experiment could settle this: have human viewers rate streamed 3DGS clips across the four network conditions on a standard opinion scale, then regress those ratings against the QoE values produced by Eqs. (38)-(42). If the QoE model does not rank-order the human ratings with a clearly positive correlation, the agent is optimizing a reward that does not match user experience, and the reported QoE superiority would not carry over to real viewers.

Watch

Extended reading notes

Core claim

The paper argues that the bottleneck in 3DGS video is transmission, not reconstruction: existing dynamic-3DGS studies build streamable sources, while tiling and bitrate control for those sources remain underdeveloped. Its central claim is that a single pipeline solves both. First, a spatiotemporal saliency network scores fine-grained tiles from Gaussian attributes (position, zero-order spherical-harmonic color, opacity, and effective volume), and agglomerative clustering merges tiles of similar saliency into irregular viewport-aligned regions. Second, each aggregated tile is classified by motion magnitude as static, low-dynamic, or high-dynamic; low-dynamic tiles share deformation-field MLPs, high-dynamic tiles get dedicated ones, and multi-resolution hash-grid features predict positional and rotational updates so client-side reconstruction only needs key frames plus deformation fields. Third, five quality tiers per tile are created by saliency-weighted Gaussian pruning, and a meta-reinforcement-learning agent chooses each tile's transmission mode (encoded versus pre-reconstructed) and quality level to maximize a layered QoE that combines peak-signal-to-noise-ratio (PSNR) weighted geometric quality, structural-similarity (SSIM) weighted rendering quality, stall time, stall frequency, and inter-GoF smoothness. Reported experiments show higher QoE and field-of-view (FoV) matching than uniform tiling and an adaptive-tiling baseline, higher QoE than four ABR baselines under 4G and 5G traces, and strong few-shot retention (84.9% of full-data performance with 20% of the training set).

Load-bearing premise

The load-bearing premise is that standard screen-quality scores computed from rendered 2D frames track what viewers actually experience in 3DGS video, even though the paper itself lists 3DGS-specific distortions—Gaussian overlap artifacts and spherical-harmonic discontinuities—that conventional metrics cannot quantify.

Editorial extensions

If this is right

  • Streaming 3DGS video would become feasible over ordinary 4G and 5G connections instead of dedicated high-bandwidth links.
  • Saliency-adaptive tiling would raise the fraction of transmitted data inside the user's field of view, cutting wasted bandwidth on unseen regions.
  • Rendering-weight-aware pruning would let a single encoded source serve multiple quality tiers, so the same content adapts to different bandwidths and devices.
  • The meta-RL controller would keep most of its performance on tiny training sets, allowing deployment in new network or content environments with little data.
  • Choosing between encoded and pre-reconstructed tiles per tile would remove the all-or-nothing tradeoff between decoding delay and bandwidth demand.

Reading between the lines

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

  • Beyond the paper, the rendering-weight criterion $w_i=\sigma_i\sqrt{\det(\Sigma_i)}$ is a natural saliency prior for other 3DGS operations such as progressive loading, viewport culling, and Level-of-Detail selection.
  • A testable extension is to validate the QoE model against human opinion scores, since the paper itself notes that PSNR/SSIM miss 3DGS-specific artifacts and the agent optimizes that unvalidated reward.
  • The motion-class grouping invites a closed-form bitrate-allocation rule that spends bandwidth in proportion to tile motion, which the paper leaves to the learned policy.
  • The task-embedding and FiLM conditioning could also adapt to client compute such as available CPU cores, which the paper keeps fixed as a constant rather than a decision variable.
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

6 major / 5 minor

Summary. This paper proposes an end-to-end streaming system for dynamic 3D Gaussian Splatting video. The server partitions each GoF into fine-grained tiles, samples Gaussians by rendering weight, extracts spatial and temporal saliency via a learned model, and clusters tiles adaptively. Tiles are encoded with deformation fields categorized by motion into static, low-dynamic, and high-dynamic classes, and quality tiers are generated by saliency-weighted Gaussian pruning. A QoE model (Eqs. 38-43) combines PSNR/SSIM of rendered 2D frames, stall time/frequency, occlusion, and saliency weights. A meta-RL ABR agent selects transmission mode and quality level per tile, trained with MAML across bandwidth and content tasks. Experiments compare sampling methods, tiling methods, and ABR baselines under 4G/5G traces and report QoE and FoV-matching gains.

Significance. If validated, the work would be a useful step toward practical 3DGS streaming, and the system-level integration of adaptive tiling, deformation-based encoding, quality tiering, and meta-RL ABR is timely and not present in prior work. The use of multiple datasets, real 4G/5G traces, and a 50-participant VR viewport study gives the evaluation a realistic basis. However, the central claims currently rest on an unvalidated QoE proxy and a partially self-referential saliency ground truth, and the experimental figures lack error bars and significance tests. The source code is promised but not provided, which also limits independent verification at this stage.

major comments (6)
  1. [Section IV-C, Eqs. (38)-(42), and Figures 5-7] The QoE metric used both as the RL reward and as the evaluation outcome is a weighted sum of PSNR and SSIM of rendered 2D frames. The Introduction states that 3DGS-specific distortions such as Gaussian overlap artifacts and spherical harmonic discontinuities 'cannot be effectively quantified by conventional quality metrics.' No term in Eqs. (38)-(42) measures representation-level degradation, so the optimization target is an unvalidated proxy. The authors should validate the QoE model against human subjective ratings (e.g., MOS) or a 3DGS-specific quality benchmark, and report the correlation, before claiming improved user experience.
  2. [Section IV-A, Eq. (26), and Section V-B, Figure 6] The ground-truth saliency score Score_j^{(t)} is computed 'as in [17]', and [17] is the same adaptive-tiling baseline (AT) against which the proposed tiling is compared in Figure 6. The tiling comparison is therefore partially self-referential. The authors should train the saliency detector with independent human fixation data or an independent saliency model, or report the correlation between the learned saliency scores and human saliency, and then re-run the comparison.
  3. [Section IV-B, Eq. (37) and surrounding text] The pruning-rate description is internally inconsistent. With p_base = 15% and alpha in [0,1], p_adj ranges over [0,15%], yet the text states the scheme produces 30-50% pruning at low quality, that the lowest quality level maintains 'at most 50% pruning', and that overall pruning remains 'below 30% at the lowest quality level'. These statements cannot all hold simultaneously. Please correct the definitions of p_base and alpha, state the actual per-level pruning rates used in the experiments, and reconcile the numbers.
  4. [Eq. (18)] The temporal saliency score O_s is written as 1 / (1 + exp(S_sim) + 1), which is not the standard sigmoid and does not lie in [0,1] as a 'saliency intensity score' should. Please correct the typo and verify that the subsequent weighting in Eq. (19) uses a properly normalized score.
  5. [Eqs. (46)-(47)] T_R^t is computed with S_{t,k,r}^E and T_E^t is computed with S_{t,k,r}^R, which is opposite to the definitions where S^E is the data size of encoded tiles and S^R is the data size of reconstructed tiles. Since these transmission times enter Eq. (48) and then the stall terms in the reward, this swap must be fixed or clearly justified if the naming is intentionally different.
  6. [Figures 5-7 and abstract] The bar charts report single values with no error bars, no standard deviations across repeated runs or datasets, and no statistical tests. The abstract's claim of 'significantly outperform' is not supported by the reported evidence. Please add confidence intervals or error bars, specify the number of seeds, and perform appropriate significance tests for the QoE and FoV-matching comparisons.
minor comments (5)
  1. [Section IV-C] The sentence 'To address the limitations of existing QoE models in evaluating 3DGS video...' is repeated verbatim twice immediately before Eq. (38); one occurrence should be removed.
  2. [Section V-A] The dataset name 'DNA-Randering' should be 'DNA-Rendering'.
  3. [Eq. (44)] The notation needs parentheses to make it clear that the entire double sum is divided by C; also, the decoding-time parameter phi_{t,k,r} should be defined more precisely.
  4. [End of Section IV-C] The text says the QoE weights lambda, mu, sigma, eta are generated by MLP(z_T), but Eqs. (38) and (53) treat them as constants; please clarify how the dynamically generated weights enter the reward.
  5. [Section V-B] The 'FoV matching rate' in Figure 6b is not defined; please state its computation (e.g., intersection over union between predicted and actual viewport regions).

Circularity Check

1 steps flagged · score 4.0 of 10

Saliency ground truth is defined by the same prior work used as the tiling baseline, making the tiling comparison partially self-referential.

  1. self citation load bearing [Section IV-A, Eq. (26)-(27); Section V-B 'Tiling method']
    "The ground truth Score(t)j is computed from static saliency detection and dynamic motion estimation as in [17]. ... we compare our method with the adaptive tiling approach (AT) proposed in [17]."

    The saliency detector is trained by minimizing the SmoothL1 loss in Eq. (26) between predicted S_j^(t) and Score_j^(t), and the ground-truth scores are defined by the same prior work [17] that later serves as the AT baseline in the tiling comparison. Thus the saliency signal is not an externally validated ground truth; it is the baseline method's own output relabeled as ground truth. Any FoV/QoE advantage of the proposed tiling is partly a comparison of a learned function approximator against its own training target. The central claim that the proposed tiling 'significantly outperform[s] state-of-the-art methods' therefore rests on a self-referential training/evaluation loop, though the ABR and encoding contributions remain independent.

full rationale

The paper is not globally circular: the meta-RL ABR agent optimizes an explicit QoE reward and is compared with external baselines, and the encoding/sampling experiments use standard PSNR plus a separate temporal-coherence metric. However, the saliency-driven tiling chain contains a load-bearing self-citation. Eq. (26) trains the saliency model against ground truth 'computed ... as in [17]', where [17] is the same group's earlier hybrid-saliency tiling method and is also the AT baseline in the tiling experiments. The reported superiority over AT is therefore partly a fit: the learner is trained to reproduce the baseline's saliency scores, so the comparison is not against independent ground truth. This is not a full 8-10 circularity because temporal saliency features, clustering, and the QoE-based evaluation could still contribute independently, and the ABR/meta-learning results are not reduced to their inputs by construction. The score of 4 reflects one self-citation that is partially load-bearing in the tiling claim while the rest of the system retains independent content.

Assumptions & free parameters 7 free parameters · 6 assumptions · 0 invented entities

The central performance claims rely on several unvalidated modeling choices: PSNR/SSIM as 3DGS quality metrics, saliency ground truth inherited from the authors' own prior work, hand-set pruning and QoE weights, and simulation-based evaluation. None of these is independently anchored to user studies or real deployments.

free parameters (7)
  • QoE weighting parameters lambda, mu, sigma, alpha (Eq. 38) = not specified
    These weights determine how quality, stalls, and stall frequency trade off in the objective that the ABR agent maximizes. They are never fit to user data or reported, so the QoE scale is arbitrary.
  • Base pruning rate p_base and pruning coefficient alpha (Eq. 37) = p_base = 15%, alpha derived but not specified
    p_base is chosen by hand from prior empirical claims, and alpha is described as derived inversely to satisfy an 8% minimum pruning constraint, but the derivation and value are omitted. The text is also internally inconsistent about whether pruning reaches 30% or 50%.
  • QoE spatial factors gamma (Eq. 40) and beta (Eq. 43) = not specified
    gamma scales salience weights in the spatial influence factor and beta controls distance attenuation in occlusion weighting; neither value is given, yet both shape the QoE reward.
  • Reward smoothness weights eta and delta (Eq. 53-54) = not specified
    These balance temporal smoothness against quality and stalls in the training reward; no values or adaptation schedule are provided.
  • Target cluster count for adaptive tiling = not specified
    The agglomerative clustering merges tiles until a predetermined threshold, but the threshold is never stated, so the resulting tile configuration is not reproducible.
  • Motion classification thresholds (static/low/high dynamic) = not specified
    Tiles are categorized by displacement magnitude thresholds, but the thresholds are not given, making the deformation-field assignment irreproducible.
  • Meta-RL learning rates and KL divergence weight xi (Eq. 55) = not specified
    The MAML inner/outer loop update equations use hyperparameters that are not reported; without them the meta-training cannot be reproduced.
assumptions (6)
  • domain assumption Conventional PSNR and SSIM are valid perceptual quality measures for rendered 3DGS video tiles.
    Used in Eq. 39 and Eq. 42 to define Qgeo and Qrender, despite the introduction stating that 3DGS-specific distortions cannot be quantified by conventional quality metrics. If PSNR/SSIM misrepresent visual quality, the whole QoE optimization is misaligned.
  • domain assumption Saliency ground-truth scores from [17] are valid for 3DGS video content.
    Eq. 26 trains the saliency network against Score_t^j 'computed ... as in [17]', transferring a point-cloud/volumetric saliency heuristic to 3DGS without validation.
  • domain assumption The rendering weight w_i = sigma_i * sqrt(det Sigma_i) measures each Gaussian's visual importance for quality-preserving pruning.
    Equations 1-2 and 34-35 use this weight for sampling and pruning. The mapping to human perception is assumed, not established.
  • domain assumption 3DGStream-style deformation fields reconstruct target frames with sufficient fidelity for streaming.
    Section IV-B builds tile deformation fields following [12]; the 33ms decode time and reconstructed quality are taken from that prior system.
  • domain assumption The simulated bandwidth, viewport, and latency model (Eqs. 44-52) accurately represent real streaming behavior.
    All QoE comparisons are computed from this simulation; no end-to-end deployment or real-time test is reported.
  • domain assumption MAML meta-learning generalizes from training bandwidth tasks to unseen network scenarios.
    Central to the claim of cross-environment generalization (Section IV-C, Eq. 55-57); no test-time adaptation evidence beyond simulated traces is provided.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Adaptive 3D Gaussian Splatting Video Streaming: Visual Saliency-Aware Tiling and Meta-Learning-Based Bitrate Adaptation." pith.science (2026). https://pith.science/paper/NVXVAPKN

@misc{pith2026250714454,
  author       = {Pith},
  title        = {Pith review of: Adaptive 3D Gaussian Splatting Video Streaming: Visual Saliency-Aware Tiling and Meta-Learning-Based Bitrate Adaptation},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/NVXVAPKN}},
  note         = {Machine review of arXiv:2507.14454}
}
read the original abstract

3D Gaussian splatting video (3DGS) streaming has recently emerged as a research hotspot in both academia and industry, owing to its impressive ability to deliver immersive 3D video experiences. However, research in this area is still in its early stages, and several fundamental challenges, such as tiling, quality assessment, and bitrate adaptation, require further investigation. In this paper, we tackle these challenges by proposing a comprehensive set of solutions. Specifically, we propose an adaptive 3DGS tiling technique guided by saliency analysis, which integrates both spatial and temporal features. Each tile is encoded into versions possessing dedicated deformation fields and multiple quality levels for adaptive selection. We also introduce a novel quality assessment framework for 3DGS video that jointly evaluates spatial-domain degradation in 3DGS representations during streaming and the quality of the resulting 2D rendered images. Additionally, we develop a meta-learning-based adaptive bitrate algorithm specifically tailored for 3DGS video streaming, achieving optimal performance across varying network conditions. Extensive experiments demonstrate that our proposed approaches significantly outperform state-of-the-art methods.

Figures

Figures reproduced from arXiv: 2507.14454 by the authors.

Figure 1
Figure 1. Server-side architecture for 3DGS video streaming. [PITH_FULL_IMAGE:figures/full_fig_p004_1.png] view at source ↗
Figure 2
Figure 2. Saliency extraction and adaptive tiling. [PITH_FULL_IMAGE:figures/full_fig_p006_2.png] view at source ↗
Figure 3
Figure 3. Quality level partitioning. The adjusted pruning rate per tile is computed as: padj = pbase ·  1 − S˜(t) m · α  , (37) where pbase denotes the base pruning rate and α ∈ [0, 1] regulates minimum pruning. Empirical results show that prun￾ing 30% of Gaussians via psample (i) preserves high visual quality [14], whereas exceeding 50% causes severe degra￾dation [15]. We therefore implement a base pruning rate of 15% per… view at source ↗
Figures from the paper (3 more)
Figure 4
Figure 4. Figure 4: Partial sequence demonstrations from our dataset. [PITH_FULL_IMAGE:figures/full_fig_p011_4.png]
Figure 6
Figure 6. Figure 6: Comparison of different tiling methods. Std4G Ext4G Std5G Ext5G Network Scenarios 0.0 0.5 1.0 1.5 2.0 2.5 3.0 3.5 4.0 4.5 5.0 QoE Values QoE Comparison for Different ABR Algorithm using Hybrid Training Dataset Algorithm Ours POT QDA HST RLA (a) Std4G Ext4G Std5G Ext5G …
Figure 7
Figure 7. Figure 7: Comparison of different ABR algorithms. temporal coherence compared to alternatives, attributable to the inherent characteristics of 3DGS. Unlike point clouds that maintain geometrically consistent distributions, 3DGS primi￾tives exhibit stochastic spatial arrangements…

Discussion (0). Sign in to comment.

Forward citations

Cited by 1 Pith paper

Reviewed papers in the Pith corpus that reference this work. Sorted by Pith novelty score. Full citation record

  1. SwiftGS: Episodic Priors for Immediate Satellite Surface Recovery

    cs.CV 2026-03 unverdicted novelty 6.0 of 10

    SwiftGS uses episodic meta-training to predict geometry-radiation-decoupled Gaussian primitives and a lightweight SDF for zero-shot 3D satellite surface reconstruction with physics-aware rendering.

Reference graph

Works this paper leans on

47 extracted references · 33 canonical work pages · cited by 1 Pith paper

  1. [17]

    Optimal volumetric video streaming with hybrid saliency based tiling,

    J. Li, C. Zhang, Z. Liu, R. Hong, and H. Hu, “Optimal volumetric video streaming with hybrid saliency based tiling,”IEEE Transactions on Multimedia, 2022

    work page 2022

  2. [1]

    Augmented and virtual reality in surgery,

    M. R. Desselle, R. A. Brown, A. R. James, M. J. Midwinter, S. K. Powell, and M. A. Woodruff, “Augmented and virtual reality in surgery,” Computing in Science & Engineering, vol. 22, no. 3, pp. 18–26, 2020

    work page 2020

  3. [2]

    Point cloud video streaming: Challenges and solutions,

    Z. Liu, Q. Li, X. Chen, C. Wu, S. Ishihara, J. Li, and Y . Ji, “Point cloud video streaming: Challenges and solutions,”IEEE Network, vol. 35, no. 5, pp. 202–209, 2021

    work page 2021

  4. [3]

    A pipeline for multiparty volumetric video conferencing: transmission of point clouds over low latency dash,

    J. Jansen, S. Subramanyam, R. Bouqueau, G. Cernigliaro, M. M. Cabr ´e, F. P ´erez, and P. Cesar, “A pipeline for multiparty volumetric video conferencing: transmission of point clouds over low latency dash,” in Proceedings of the 11th ACM Multimedia Systems Conference, 2020, pp. 341–344

    work page 2020

  5. [4]

    From capture to display: A survey on volumetric video,

    Y . Jin, K. Hu, J. Liu, F. Wang, and X. Liu, “From capture to display: A survey on volumetric video,”arXiv preprint arXiv:2309.05658, 2023

    arXiv 2023

  6. [5]

    Nerf- editing: geometry editing of neural radiance fields,

    Y .-J. Yuan, Y .-T. Sun, Y .-K. Lai, Y . Ma, R. Jia, and L. Gao, “Nerf- editing: geometry editing of neural radiance fields,” inProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2022, pp. 18 353–18 364

    work page 2022

  7. [6]

    Gaussian grouping: Segment and edit anything in 3d scenes,

    M. Ye, M. Danelljan, F. Yu, and L. Ke, “Gaussian grouping: Segment and edit anything in 3d scenes,” inEuropean Conference on Computer Vision. Springer, 2024, pp. 162–179

    work page 2024

  8. [7]

    Robust dual gaussian splatting for immersive human-centric volumetric videos,

    Y . Jiang, Z. Shen, Y . Hong, C. Guo, Y . Wu, Y . Zhang, J. Yu, and L. Xu, “Robust dual gaussian splatting for immersive human-centric volumetric videos,”ACM Transactions on Graphics (TOG), vol. 43, no. 6, pp. 1–15, 2024

    work page 2024

Show all 47 references
  1. [8]

    Lts: A dash streaming system for dynamic multi-layer 3d gaussian splatting scenes,

    Y .-C. Sun, Y . Shi, C.-T. Lee, M. Zhu, W. T. Ooi, Y . Liu, C.-Y . Huang, and C.-H. Hsu, “Lts: A dash streaming system for dynamic multi-layer 3d gaussian splatting scenes,” inProceedings of the 16th ACM Multimedia Systems Conference, 2025, pp. 136–147

    2025

  2. [9]

    Recent advances in 3d gaussian splatting,

    T. Wu, Y .-J. Yuan, L.-X. Zhang, J. Yang, Y .-P. Cao, L.-Q. Yan, and L. Gao, “Recent advances in 3d gaussian splatting,”Computational Visual Media, pp. 1–30, 2024

    2024

  3. [10]

    Compact 3d gaussian representation for radiance field,

    J. C. Lee, D. Rho, X. Sun, J. H. Ko, and E. Park, “Compact 3d gaussian representation for radiance field,” inProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2024, pp. 21 719–21 728

    2024

  4. [11]

    Gs-qa: Com- prehensive quality assessment benchmark for gaussian splatting view synthesis,

    P. Martin, A. Rodrigues, J. Ascenso, and M. P. Queluz, “Gs-qa: Com- prehensive quality assessment benchmark for gaussian splatting view synthesis,”arXiv preprint arXiv:2502.13196, 2025

    2025 arXiv

  5. [12]

    3dgstream: On- the-fly training of 3d gaussians for efficient streaming of photo-realistic free-viewpoint videos,

    J. Sun, H. Jiao, G. Li, Z. Zhang, L. Zhao, and W. Xing, “3dgstream: On- the-fly training of 3d gaussians for efficient streaming of photo-realistic free-viewpoint videos,” inProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2024, pp. 20 675–20 685

    2024

  6. [13]

    Multi- frame bitrate allocation of dynamic 3d gaussian splatting streaming over dynamic networks,

    Y .-C. Sun, Y . Shi, W. T. Ooi, C.-Y . Huang, and C.-H. Hsu, “Multi- frame bitrate allocation of dynamic 3d gaussian splatting streaming over dynamic networks,” inProceedings of the 2024 SIGCOMM Workshop on Emerging Multimedia Systems, 2024, pp. 1–7

    2024

  7. [14]

    Vˆ 3: Viewing volumetric videos on mobiles via streamable 2d dynamic gaussians,

    P. Wang, Z. Zhang, L. Wang, K. Yao, S. Xie, J. Yu, M. Wu, and L. Xu, “Vˆ 3: Viewing volumetric videos on mobiles via streamable 2d dynamic gaussians,”ACM Transactions on Graphics (TOG), vol. 43, no. 6, pp. 1–13, 2024

    2024

  8. [15]

    L3gs: Layered 3d gaussian splats for efficient 3d scene delivery,

    Y .-Z. Tsai, X. Zhang, Z. Li, and J. Chen, “L3gs: Layered 3d gaussian splats for efficient 3d scene delivery,”arXiv preprint arXiv:2504.05517, 2025

    2025 arXiv

  9. [16]

    Pcgs: Progressive compression of 3d gaussian splatting,

    Y . Chen, M. Li, Q. Wu, W. Lin, M. Harandi, and J. Cai, “Pcgs: Progressive compression of 3d gaussian splatting,”arXiv preprint arXiv:2503.08511, 2025

    2025 arXiv

  10. [18]

    Metaabr: A meta-learning approach on adaptative bitrate selection for video streaming,

    W. Li, X. Li, Y . Xu, Y . Yang, and S. Lu, “Metaabr: A meta-learning approach on adaptative bitrate selection for video streaming,”IEEE Transactions on Mobile Computing, vol. 23, no. 3, pp. 2422–2437, 2023

    2023

  11. [19]

    Bitrate adaptation and guidance with meta reinforcement learning,

    A. Bentaleb, M. Lim, M. N. Akcay, A. C. Begen, and R. Zimmermann, “Bitrate adaptation and guidance with meta reinforcement learning,” IEEE Transactions on Mobile Computing, 2024

    2024

  12. [20]

    Dynamic scene reconstruction: Recent advance in real-time rendering and streaming,

    J. Zhu and H. Tang, “Dynamic scene reconstruction: Recent advance in real-time rendering and streaming,”arXiv preprint arXiv:2503.08166, 2025

    2025 arXiv

  13. [21]

    Streamgs: Online generalizable gaussian splatting reconstruction for unposed image streams,

    Y . Li, J. Wang, L. Chu, X. Li, S.-h. Kao, Y .-C. Chen, and Y . Lu, “Streamgs: Online generalizable gaussian splatting reconstruction for unposed image streams,”arXiv preprint arXiv:2503.06235, 2025

    2025 arXiv

  14. [22]

    Repre- senting long volumetric video with temporal gaussian hierarchy,

    Z. Xu, Y . Xu, Z. Yu, S. Peng, J. Sun, H. Bao, and X. Zhou, “Repre- senting long volumetric video with temporal gaussian hierarchy,”ACM Transactions on Graphics (TOG), vol. 43, no. 6, pp. 1–18, 2024

    2024

  15. [23]

    Instant gaussian stream: Fast and generalizable streaming of dynamic scene reconstruction via gaussian splatting,

    J. Yan, R. Peng, Z. Wang, L. Tang, J. Yang, J. Liang, J. Wu, and R. Wang, “Instant gaussian stream: Fast and generalizable streaming of dynamic scene reconstruction via gaussian splatting,”arXiv preprint arXiv:2503.16979, 2025

    2025 arXiv

  16. [24]

    Evolvinggs: High-fidelity streamable volumetric video via evolving 3d gaussian representation,

    C. Zhang, Y . Zhou, S. Wang, W. Li, D. Wang, Y . Xu, and S. Jiao, “Evolvinggs: High-fidelity streamable volumetric video via evolving 3d gaussian representation,”arXiv preprint arXiv:2503.05162, 2025. 14

    2025 arXiv

  17. [25]

    Qoe-driven adaptive streaming for point clouds,

    L. Wang, C. Li, W. Dai, S. Li, J. Zou, and H. Xiong, “Qoe-driven adaptive streaming for point clouds,”IEEE Transactions on Multimedia, 2022

    2022

  18. [26]

    Toward optimal real-time volumetric video streaming: A rolling optimization and deep reinforcement learning based approach,

    J. Li, H. Wang, Z. Liu, P. Zhou, X. Chen, Q. Li, and R. Hong, “Toward optimal real-time volumetric video streaming: A rolling optimization and deep reinforcement learning based approach,”IEEE Transactions on Circuits and Systems for Video Technology, vol. 33, no. 12, pp. 7870–...

    2023

  19. [27]

    Efficient volumetric video streaming through super resolution,

    A. Zhang, C. Wang, B. Han, and F. Qian, “Efficient volumetric video streaming through super resolution,” inProceedings of the 22nd Interna- tional Workshop on Mobile Computing Systems and Applications, 2021, pp. 106–111

    2021

  20. [28]

    Toward holographic video communications: a promising ai-driven solution,

    Y . Huang, Y . Zhu, X. Qiao, X. Su, S. Dustdar, and P. Zhang, “Toward holographic video communications: a promising ai-driven solution,” IEEE Communications Magazine, vol. 60, no. 11, pp. 82–88, 2022

    2022

  21. [29]

    Cav3: Cache-assisted viewport adaptive volumetric video streaming,

    J. Liu, B. Zhu, F. Wang, Y . Jin, W. Zhang, Z. Xu, and S. Cui, “Cav3: Cache-assisted viewport adaptive volumetric video streaming,” in2023 IEEE Conference Virtual Reality and 3D User Interfaces (VR). IEEE, 2023, pp. 173–183

    2023

  22. [30]

    Progressive frame patching for fov-based point cloud video streaming,

    T. Zong, Y . Mao, C. Li, Y . Liu, and Y . Wang, “Progressive frame patching for fov-based point cloud video streaming,”IEEE Transactions on Multimedia, 2025

    2025

  23. [31]

    Qv4: Qoe-based viewpoint-aware v-pcc-encoded volumetric video streaming,

    Y . Shi, B. Clement, and W. T. Ooi, “Qv4: Qoe-based viewpoint-aware v-pcc-encoded volumetric video streaming,” inProceedings of the 15th ACM Multimedia Systems Conference, 2024, pp. 144–154

    2024

  24. [32]

    Towards full-scene volumetric video streaming via spatially layered representation and nerf generation,

    J. Shi, M. Zhang, L. Shen, J. Liu, Y . Zhang, L. Pu, and J. Xu, “Towards full-scene volumetric video streaming via spatially layered representation and nerf generation,” inProceedings of the 34th edition of the Workshop on Network and Operating System Support for Digital Audio...

    2024

  25. [33]

    Lapisgs: Layered progressive 3d gaussian splatting for adaptive streaming,

    Y . Shi, G. Morin, S. Gasparini, and W. T. Ooi, “Lapisgs: Layered progressive 3d gaussian splatting for adaptive streaming,”arXiv preprint arXiv:2408.14823, 2024

    2024

  26. [34]

    Trend analysis using agglomerative hierarchical clustering approach for time series big data,

    S. Pasupathi, V . Shanmuganathan, K. Madasamy, H. R. Yesudhas, and M. Kim, “Trend analysis using agglomerative hierarchical clustering approach for time series big data,”The Journal of Supercomputing, vol. 77, no. 7, pp. 6505–6524, 2021

    2021

  27. [35]

    Emerging mpeg standards for point cloud compression,

    S. Schwarz, M. Preda, V . Baroncini, M. Budagavi, P. Cesar, P. A. Chou, R. A. Cohen, M. Krivoku ´ca, S. Lasserre, Z. Liet al., “Emerging mpeg standards for point cloud compression,”IEEE Journal on Emerging and Selected Topics in Circuits and Systems, vol. 9, no. 1, pp. 133–148, 2018

    2018

  28. [36]

    Hifi4g: High-fidelity human performance rendering via compact gaussian splatting,

    Y . Jiang, Z. Shen, P. Wang, Z. Su, Y . Hong, Y . Zhang, J. Yu, and L. Xu, “Hifi4g: High-fidelity human performance rendering via compact gaussian splatting,” inProceedings of the IEEE/CVF conference on computer vision and pattern recognition, 2024, pp. 19 734–19 745

    2024

  29. [37]

    Dna-rendering: A diverse neural actor repository for high-fidelity human-centric rendering,

    W. Cheng, R. Chen, S. Fan, W. Yin, K. Chen, Z. Cai, J. Wang, Y . Gao, Z. Yu, Z. Linet al., “Dna-rendering: A diverse neural actor repository for high-fidelity human-centric rendering,” inProceedings of the IEEE/CVF International Conference on Computer Vision, 2023, pp. 19 982–19 993

    2023

  30. [38]

    Vpformer: Leveraging transformer with voxel integration for viewport prediction in volumetric video,

    J. Li, Z. Zhao, Q. Li, Z. Li, P. Y . Zhou, Z. Liu, H. Zhou, and Z. Li, “Vpformer: Leveraging transformer with voxel integration for viewport prediction in volumetric video,”ACM Transactions on Multimedia Computing, Communications and Applications

  31. [39]

    Beyond throughput: A 4g lte dataset with channel and context metrics,

    D. Raca, J. J. Quinlan, A. H. Zahran, and C. J. Sreenan, “Beyond throughput: A 4g lte dataset with channel and context metrics,” in Proceedings of the 9th ACM multimedia systems conference, 2018, pp. 460–465

    2018

  32. [40]

    Realtime mobile bandwidth prediction using lstm neural network and bayesian fusion,

    L. Mei, R. Hu, H. Cao, Y . Liu, Z. Han, F. Li, and J. Li, “Realtime mobile bandwidth prediction using lstm neural network and bayesian fusion,”Computer Networks, vol. 182, p. 107515, 2020

    2020

  33. [41]

    Beyond throughput, the next generation: A 5g dataset with channel and context metrics,

    D. Raca, D. Leahy, C. J. Sreenan, and J. J. Quinlan, “Beyond throughput, the next generation: A 5g dataset with channel and context metrics,” in Proceedings of the 11th ACM multimedia systems conference, 2020, pp. 303–308

    2020

  34. [42]

    Viewport prediction for volumetric video streaming by exploring video saliency and trajectory information,

    J. Li, Z. Li, Z. Liu, P. Zhou, R. Hong, Q. Li, and H. Hu, “Viewport prediction for volumetric video streaming by exploring video saliency and trajectory information,”IEEE Transactions on Circuits and Systems for Video Technology, 2025

    2025

  35. [43]

    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,”Advances in neural information processing systems, vol. 30, 2017

    2017

  36. [44]

    V oxelnet: End-to-end learning for point cloud based 3d object detection,

    Y . Zhou and O. Tuzel, “V oxelnet: End-to-end learning for point cloud based 3d object detection,” inProceedings of the IEEE conference on computer vision and pattern recognition, 2018, pp. 4490–4499

    2018

  37. [45]

    Flex-convolution: Million- scale point-cloud learning beyond grid-worlds,

    F. Groh, P. Wieschollek, and H. P. Lensch, “Flex-convolution: Million- scale point-cloud learning beyond grid-worlds,” inAsian Conference on Computer Vision. Springer, 2018, pp. 105–122

    2018

  38. [46]

    Qoe-driven dynamic adaptive video streaming strategy with future information,

    L. Yu, T. Tillo, and J. Xiao, “Qoe-driven dynamic adaptive video streaming strategy with future information,”IEEE Transactions on Broadcasting, vol. 63, no. 3, pp. 523–534, 2017

    2017

  39. [47]

    Reinforcement learning-based adaptation and scheduling methods for multi-source dash,

    N. T. Nguyen, L. Luu, P. L. V o, S. T. T. Nguyen, C. T. Do, and N.- T. Nguyen, “Reinforcement learning-based adaptation and scheduling methods for multi-source dash,”Computer Science and Information Systems, vol. 20, no. 1, pp. 157–173, 2023

    2023

Pith tools

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