Pith. sign in

REVIEW 3 major objections 5 minor 46 references

A Novel Benchmark and Dataset for Efficient 3D Gaussian Splatting with Gaussian Point Cloud Compression

T0 review · 3 major / 5 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 A learned codec trained on 1,000 Gaussian scenes cuts the bitrate of Gaussian positions by 8.2% over MPEG G-PCC v23 and runs about six times faster.

desk verdict Useful dataset and a plausible geometry codec, but the integration results are confounded by anchor-count changes that a lossless position codec cannot explain. read the letter →

arxiv 2505.18197 v1 pith:IW624YZB submitted 2025-05-21 cs.GR

Kangli Wang , Shihao Li , Qianxi Yi , Wei Gao This is my paper · ORCID
classification cs.GR
keywords 3DGaussianSplattingcompressionpointcloudgeometrylearnedoccupancycodepredictionG-PCCGausPcc-1KdatasetentropycodingScaffold-GS
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

3D Gaussian Splatting scenes are large, and the spatial positions of Gaussians account for a substantial, usually under-compressed share of the bitstream. The paper argues that those positions form a point cloud with a distinctive geometry, globally sparse but locally dense, and that learned point-cloud codecs can compress them far better than the MPEG G-PCC standard if trained on the right data. To make that possible, it constructs GausPcc-1K, a dataset of 1,000 trained Gaussian scenes, and GausPcgc, an entropy-coded predictor of voxel occupancy. On the paper's benchmarks GausPcgc beats G-PCC v23 by 8.2% in rate while running about six times faster, and it shrinks the total bitstreams of HAC, HAC++, CAT-3DGS, and TC-GS with no loss of rendering quality. A sympathetic reader would take the paper to establish that Gaussian position compression is a separate, tractable problem with its own data distribution and its own benchmark.

What carries the argument

The object that carries the argument is the occupancy code, an eight-bit pattern that records which of a voxel's eight children are occupied at the next finer scale of a multi-scale voxel hierarchy. GausPcgc's Four-stage Occupancy Predictor (FOP) estimates the probability of each occupancy code in a 1-1-2-4 bit split, predicting bit 1, then bit 2, then bits 3-4, then bits 5-8, with each stage conditioned on the previously predicted bits and on features aggregated from neighboring voxels by sparse convolutions; the resulting cross-entropy loss is the per-point bitrate, and arithmetic coding turns the predicted distributions into a bitstream. The companion GausPcc-1K dataset supplies the training distribution: 1,000 Scaffold-GS scenes curated from DL3DV with average PSNR 29.01 dB, chosen to reproduce the locally dense, globally sparse geometry of Gaussian point clouds.

What would settle it

Train GausPcgc on GausPcc-1K and measure bits per point separately for every Mip-NeRF360, Tanks and Temples, and Deep Blending scene, not just the pooled average; if individual scenes land far from the 13.27 bpp mean, or if any scene's local-density histogram is closer to Kitti's than to GausPcc-1K's, the dataset-transfer premise behind the 8.2% gain fails. The paper's Table 1 already contains a version of this test: Octattention trained on Kitti scores 11.31 bpp on the same Gaussian positions, below GausPcgc's 13.27 bpp, so the practical claim stands only when decode latency is part of the comparison.

Watch

Extended reading notes

Core claim

The paper's central claim is that Gaussian positions can be compressed losslessly and far more efficiently by a learned geometry codec trained specifically on Gaussian-like point clouds, and that this codec can be dropped into existing 3DGS compression pipelines. Concretely, GausPcgc converts each Gaussian or anchor point set into multi-scale voxel occupancy codes and predicts each eight-bit code in four conditioned stages (bit 1, then bit 2, then bits 3-4, then bits 5-8), using sparse 3D convolutions over neighboring voxels; the predicted probabilities drive arithmetic coding. Trained on GausPcc-1K, it reaches 13.27 bpp average on the Mip-NeRF360, Tanks and Temples, and Deep Blending test positions versus 14.46 bpp for G-PCC v23, an 8.2% rate gain, with 0.797 s encoding and 0.834 s decoding versus 5.72 s and 4.04 s for G-PCC. Inserted into HAC, HAC++, CAT-3DGS, and TC-GS, it reduces total model size while holding or slightly improving PSNR; on the HAC bicycle scene the total bitstream drops from 45.67 MB to 36.39 MB, with the position stream falling to 1.08 MB and PSNR rising from 25.01 to 25.14 dB.

Load-bearing premise

The load-bearing premise is that the 1,000 Scaffold-GS scenes in GausPcc-1K, curated from DL3DV, share the position distribution of the Gaussian test scenes from Mip-NeRF360, Tanks and Temples, and Deep Blending, so a codec trained on the dataset transfers to those scenes; the paper validates this only with aggregate KL divergences, not per-scene or per-method transfer, and that premise carries the 8.2% gain in Table 1.

Editorial extensions

If this is right

  • Gaussian position storage drops from the common 48 bits per point (3 coordinates times 16-bit quantization) to roughly 13.27 bits per point on the paper's test sets, so total 3DGS model sizes shrink without retraining the scene.
  • Swapping G-PCC for GausPcgc inside HAC, HAC++, CAT-3DGS, and TC-GS reduces total bitstream size while preserving or slightly improving PSNR, SSIM, and LPIPS in the reported scenes.
  • Because the position codec is lossless relative to voxel-size quantization, it avoids the duplicate-coordinate artifacts the paper attributes to 16-bit position quantization in HAC and CAT-3DGS.
  • The benchmark establishes that point-cloud codecs trained on GausPcc-1K, rather than on Kitti or 8iVFB, are the ones that beat G-PCC on Gaussian geometry, giving future work a standard training set.
  • GausPcc-1K retains attribute values alongside geometry, so the same dataset can later support learned Gaussian attribute compression, the paper's stated future direction.

Reading between the lines

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

  • If the distribution-transfer premise holds, the same recipe, gather a few hundred trained scenes from a target representation, fit a learned occupancy predictor, and plug it into the existing compressor, should transfer to other anchor-based Gaussian variants, not just Scaffold-GS.
  • The paper's own Table 1 shows that a Kitti-trained Octattention model reaches 11.31 bpp, below GausPcgc's 13.27 bpp, so in a regime where decode time is irrelevant the best-rate claim would not hold; GausPcgc's practical contribution is a rate-latency trade-off, not rate supremacy.
  • Because GausPcc-1K is generated from Scaffold-GS anchors, a user working with dense per-Gaussian 3DGS rather than structured anchors may need a new training set, a regime the paper's benchmark does not directly test.
  • A testable extension is to compare GausPcgc against a model trained on Kitti plus a small fine-tuning set of Gaussian scenes; if fine-tuning closes the gap, dataset size rather than distribution may be the active ingredient.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

3 major / 5 minor

Summary. The paper introduces GausPcc-1K, a dataset of 1,000 Scaffold-GS Gaussian point clouds curated from DL3DV, and GausPcgc, an AI-based lossless geometry codec for Gaussian positions that uses hierarchical occupancy codes and a four-stage conditional probability model. The authors report an 8.2% bits-per-point reduction over G-PCC v23 with roughly 6x faster inference on a new Gaussian point cloud compression benchmark, and they claim that plugging GausPcgc into HAC, HAC++, CAT-3DGS, and TC-GS reduces total bitstream size without visual quality loss. The paper also presents distribution analyses (local density, fractal dimension, KL divergence) to justify the need for a specialized dataset.

Significance. If the claims were supported by consistent evidence, the contribution would be valuable: the dataset creation effort is substantial, the distribution analysis is a useful step for bridging point cloud compression and 3DGS, and an AI codec that beats G-PCC v23 on Gaussian positions with lower latency would have practical impact. The authors commit to releasing code, data, and models, which would aid reproducibility. However, the current manuscript contains internal inconsistencies in the central benchmark table and uncontrolled integration experiments, so the headline claims are not yet established. The novelty of the dataset and the integration idea are genuine, but the evidence presented is not currently proportionate to the strength of the claims.

major comments (3)
  1. [§5.2, Table 1] The text states that “Models trained on Kitti fail to outperform G-PCC” and that retraining on GausPcc-1K makes “most methods surpass G-PCC,” but Table 1 contradicts both statements. Kitti-trained Octattention achieves 11.31 bpp versus G-PCC’s 14.46 bpp, and Kitti-trained EHEM (13.37) and RENO (13.89) also beat G-PCC. Meanwhile, among GausPcc-trained models, only RENO (Ours) and the proposed Ours column beat G-PCC, while SparsePCGC (Ours), Octattention (Ours), and EHEM (Ours) do not. Furthermore, the RENO (Ours) column is numerically identical to the proposed Ours column in every row, including Avg Bpp (13.27) and CR Gain (−8.2%), which indicates a table construction error. This table is the core evidence for the benchmark and for the claimed 8.2% gain, and it must be corrected and re-analyzed before the central claims can be assessed.
  2. [§5.2, Tables 3, 6, 7 and Appendix B.4/B.5] The integration experiments are confounded by differing anchor counts: Ours-HAC has 819,267 anchors versus HAC’s 912,838 (10.2% fewer), Ours-Cat-3DGS has 601,928 versus 623,483 (3.5% fewer), and Ours-TC-GS has 450,765 versus 511,792 (11.9% fewer). Appendix B.4 states that GausPcgc uses voxel-size-based quantization that guarantees “strictly lossless geometric representation,” but a lossless position codec cannot change the number of anchors in a trained model. Appendix B.5 then attributes the reduced anchor count to the compression method, which is internally inconsistent. Consequently, the size reductions and PSNR changes in Table 2 cannot be attributed to GausPcgc: a large portion of the savings comes from the reduced anchor count, as seen in Table 3 where feature, scaling, and offset costs fall by 5.12 MB combined while the position saving is only 4.14 MB. The authors should either run controlled experiments with matched anchor counts or explicitly state that the quantization is lossy and analyze the resulting rate-distortion trade-off.
  3. [§4.2, Table 8] The transferability of GausPcc-1K to the test scenes is a load-bearing premise for the claimed 8.2% gain, but it is supported only by aggregate KL divergences averaged across scales. No per-scene or per-method transfer analysis is provided, and the test point clouds in Section 5.1 are extracted from trained 3DGS models that likely use the same Scaffold-GS pipeline used to create GausPcc-1K. The fact that Kitti-trained Octattention already beats G-PCC on Gaussian point clouds (Table 1) further weakens the claim that the distribution mismatch is severe for all methods. The authors should provide per-scene transfer results and clarify the extent to which the evaluation set overlaps with the training distribution.
minor comments (5)
  1. [Table 1] The header layout with “Test/Train” is very difficult to parse, and the duplicated RENO/Ours columns should be corrected. Please also ensure that the CR Gain values are consistently defined (negative indicates improvement) and that the highlighted best/second-best cells match the corrected numbers.
  2. [Appendix B.1] The benchmark protocol mixes own reproductions (G-PCC, RENO on 8iVFB) with values taken from original papers’ RD curves (SparsePCGC, Octattention, EHEM). This should be disclosed in the main text, since differing reproduction protocols can bias the comparison.
  3. [§4.2] The dataset selection criteria are described only as “rigorous quality assessment criteria”; please specify the PSNR threshold or other filtering rules used to select the 1,000 scenes from the trained samples.
  4. [Appendix B.5] The sentence “our method requires fewer anchors to represent the entire scene” directly conflicts with the lossless claim in Appendix B.4. If the method is lossless with respect to voxel-size quantization, the anchor set should be unchanged; please reword or correct this statement.
  5. [Figure 5] The caption contains a typo: “llustration” should be “Illustration.”

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the central compression benchmark trains on GausPcc-1K and tests on external Gaussian scenes against external baselines; reported gains are empirical. The anchor-count changes in integration experiments are a non-circular experimental confound.

full rationale

The paper's core result—GausPcgc's bitrate on Gaussian point clouds—is obtained by training on the newly constructed GausPcc-1K dataset and evaluating on the external Mip-NeRF360, Tanks and Temples, and Deep Blending scenes, with G-PCC v23 and other point cloud codecs as external baselines. No tested quantity is fitted to the evaluation data, and no reported number is defined in terms of the quantity it is claimed to predict. The dataset-suitability argument (local density, fractal dimension, KL divergence) is an empirical comparison rather than a derivation: GausPcc-1K is built from Scaffold-GS models on DL3DV scenes, while the test distributions are measured independently, so the claimed match is not an identity by construction. The self-citations (notably UniPCGC for non-uniform grouping and Octattention as a baseline) are transparent and individually ablated; they provide components but do not carry the central generalization claim. The main weaknesses are internal-validity issues rather than circularity: Appendix B.5 reports that 'Ours-*' variants have substantially fewer anchors than the baselines (e.g., HAC 912,838 to 819,267 in Table 3), which is hard to reconcile with the paper's assertion that GausPcgc provides 'strictly lossless geometric representation' and therefore should not change the anchor count; as a result, the size reductions in Table 2 may be partly attributable to model retraining or quantization changes rather than to the position codec itself. Similarly, the absence of a direct comparison with the authors' own UniPCGC baseline is a completeness concern. These issues affect the attribution of the empirical gains, but they do not make the derivation circular: the benchmark numbers remain externally evaluated results rather than reductions to the paper's own assumptions.

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

The contribution is empirical; no new physical entities or first-principles derivations are involved. The central claims rest on hand-designed architecture choices (grouping, kernel size, channels), on the choice of Scaffold-GS anchors as the geometry to compress, and on the assumption that the GausPcc-1K training distribution matches the test distribution.

free parameters (3)
  • Non-uniform grouping structure = 1-1-2-4
    Hand-designed bit split of the 8-bit occupancy code in Section 4.3 and Appendix A.1; ablation Table 4 shows it changes bpp, so it is a tuned design choice.
  • Spatial convolution kernel size = 5
    All spatial convolutions use kernel size k=5 in Appendix A.1; chosen without formal justification, and it affects context modeling.
  • Feature channel dimension = 32
    Feature channels C=32 in Appendix A.1; a hyperparameter affecting model capacity and compression performance.
assumptions (4)
  • domain assumption Scaffold-GS structured anchors (Eq. 2) are the geometry representation to compress, and Gaussian positions in other 3DGS variants share this distribution.
    The dataset and method are built on Scaffold-GS anchor positions in Section 4.2; if raw 3DGS positions are used, the learned prior may not transfer.
  • standard math Arithmetic coding with predicted occupancy-code probabilities achieves lossless compression after quantization.
    Section 4.3 and Algorithm 1 rely on standard entropy coding results, not derived in the paper.
  • domain assumption KL divergence and fractal dimension statistics at chosen voxel scales capture the properties relevant to compression performance.
    Section 4.2 and Appendix C motivate GausPcc-1K through these statistics; the link to actual codec performance is empirical only.
  • domain assumption The chosen test scenes are representative of the target Gaussian compression domain.
    Evaluation uses Mip-NeRF360, Tanks and Temples, and Deep Blending in Section 5.1; no analysis covers how these sets span the distribution of all Gaussian scenes.

how reviews work

0 comments
Cite this review

Pith. "Pith review of A Novel Benchmark and Dataset for Efficient 3D Gaussian Splatting with Gaussian Point Cloud Compression." pith.science (2026). https://pith.science/paper/IW624YZB

@misc{pith2026250518197,
  author       = {Pith},
  title        = {Pith review of: A Novel Benchmark and Dataset for Efficient 3D Gaussian Splatting with Gaussian Point Cloud Compression},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/IW624YZB}},
  note         = {Machine review of arXiv:2505.18197}
}
read the original abstract

Recently, immersive media and autonomous driving applications have significantly advanced through 3D Gaussian Splatting (3DGS), which offers high-fidelity rendering and computational efficiency. Despite these advantages, 3DGS as a display-oriented representation requires substantial storage due to its numerous Gaussian attributes. Current compression methods have shown promising results but typically neglect the compression of Gaussian spatial positions, creating unnecessary bitstream overhead. We conceptualize Gaussian primitives as point clouds and propose leveraging point cloud compression techniques for more effective storage. AI-based point cloud compression demonstrates superior performance and faster inference compared to MPEG Geometry-based Point Cloud Compression (G-PCC). However, direct application of existing models to Gaussian compression may yield suboptimal results, as Gaussian point clouds tend to exhibit globally sparse yet locally dense geometric distributions that differ from conventional point cloud characteristics. To address these challenges, we introduce GausPcgc for Gaussian point cloud geometry compression along with a specialized training dataset GausPcc-1K. Our work pioneers the integration of AI-based point cloud compression into Gaussian compression pipelines, achieving superior compression ratios. The framework complements existing Gaussian compression methods while delivering significant performance improvements. All code, data, and pre-trained models will be publicly released to facilitate further research advances in this field.

Figures

Figures reproduced from arXiv: 2505.18197 by the authors.

Figure 1
Figure 1. Overview of our GausPcc-1K dataset and GausPcgc framework. Existing methods neglect [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. Visualization of local density. We count neighbors within a 5×5×5 vicinity and render the [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗
Figure 3
Figure 3. Comparative analysis of local density and fractal dimension across different datasets. [PITH_FULL_IMAGE:figures/full_fig_p004_3.png] view at source ↗
Figures from the paper (12 more)
Figure 4
Figure 4. Figure 4: Introduction of the proposed GausPcc-1K Dataset. [PITH_FULL_IMAGE:figures/full_fig_p005_4.png]
Figure 11
Figure 11. Figure 11: As shown, existing point clouds at certain scales (8iVFB, Kitti, ScanNet) exhibit a [PITH_FULL_IMAGE:figures/full_fig_p005_11.png]
Figure 5
Figure 5. Figure 5: llustration of the proposed GausPcgc framework. [PITH_FULL_IMAGE:figures/full_fig_p006_5.png]
Figure 6
Figure 6. Figure 6: Visual comparison on Tanks and Temples [17] train scene. [PITH_FULL_IMAGE:figures/full_fig_p009_6.png]
Figure 7
Figure 7. Figure 7: Detailed framework diagram of Four-stage Occupancy Predictor. [PITH_FULL_IMAGE:figures/full_fig_p014_7.png]
Figure 8
Figure 8. Figure 8: Visual comparison of rendering results on the Mip-NeRF 360 [ [PITH_FULL_IMAGE:figures/full_fig_p016_8.png]
Figure 9
Figure 9. Figure 9: Visual comparison of rendering results on the Tanks and Temples [ [PITH_FULL_IMAGE:figures/full_fig_p017_9.png]
Figure 10
Figure 10. Figure 10: RD curves of various methods. For better visualization, we select only the top 14 methods [PITH_FULL_IMAGE:figures/full_fig_p018_10.png]
Figure 11
Figure 11. Figure 11: Local density comparison of 8iVFB, Gaussian, GausPcc, Kitti, and ScanNet datasets [PITH_FULL_IMAGE:figures/full_fig_p020_11.png]
Figure 12
Figure 12. Figure 12: Visualization of local density for more samples. All point clouds are visualized at Scale [PITH_FULL_IMAGE:figures/full_fig_p021_12.png]
Figure 13
Figure 13. Figure 13: Fractal Dimension comparison of 8iVFB, Gaussian, GausPcc, Kitti, and ScanNet datasets [PITH_FULL_IMAGE:figures/full_fig_p021_13.png]
Figure 12
Figure 12. Figure 12: D Limitations, Future Work and Broader impacts D.1 Limitations The proposed method inevitably introduces temporal latency during encoding and decoding processes, and its encoding efficiency may be compromised when the scene contains numerous duplicate coordinate point…

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

46 extracted references · 31 canonical work pages

  1. [1]

    M. T. Bagdasarian, P. Knoll, Y .-H. Li, F. Barthel, A. Hilsmann, P. Eisert, and W. Morgenstern. 3dgs. zip: A survey on 3d gaussian splatting compression methods.arXiv preprint arXiv:2407.09510, 2024

    arXiv 2024

  2. [2]

    J. T. Barron, B. Mildenhall, D. Verbin, P. P. Srinivasan, and P. Hedman. Mip-nerf 360: Unbounded anti-aliased neural radiance fields. InProceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 5470–5479, 2022

    2022

  3. [3]

    C. Cao, M. Preda, V . Zakharchenko, E. S. Jang, and T. Zaharia. Compression of sparse and dense dynamic point clouds—methods and standards.Proceedings of the IEEE, 109(9):1537–1558, 2021

    work page 2021

  4. [4]

    A. Chen, Z. Xu, A. Geiger, J. Yu, and H. Su. Tensorf: Tensorial radiance fields. InEuropean conference on computer vision, pages 333–350. Springer, 2022

    work page 2022

  5. [5]

    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

    arXiv 2025

  6. [6]

    Y . Chen, Q. Wu, W. Lin, M. Harandi, and J. Cai. Hac: Hash-grid assisted context for 3d gaussian splatting compression. InEuropean Conference on Computer Vision, pages 422–438. Springer, 2024

    work page 2024

  7. [7]

    Y . Chen, Q. Wu, W. Lin, M. Harandi, and J. Cai. Hac++: Towards 100x compression of 3d gaussian splatting.arXiv preprint arXiv:2501.12255, 2025

    arXiv 2025

  8. [8]

    A. Dai, A. X. Chang, M. Savva, M. Halber, T. Funkhouser, and M. Nießner. Scannet: Richly-annotated 3d reconstructions of indoor scenes. InProceedings of the IEEE conference on computer vision and pattern recognition, pages 5828–5839, 2017

    work page 2017

Show all 46 references
  1. [9]

    d’Eon, H

    E. d’Eon, H. Bob, T. Myers, and P. A. Chou. 8i voxelized full bodies - a voxelized point cloud dataset. InISO/IEC JTC1/SC29 Joint WG11/WG1 (MPEG/JPEG) input document WG11M40059/WG1M74006, 2017

    2017

  2. [10]

    Z. Fan, K. Wang, K. Wen, Z. Zhu, D. Xu, Z. Wang, et al. Lightgaussian: Unbounded 3d gaussian compression with 15x reduction and 200+ fps.Advances in neural information processing systems, 37:140138–140158, 2024

    2024

  3. [11]

    Fridovich-Keil, G

    S. Fridovich-Keil, G. Meanti, F. R. Warburg, B. Recht, and A. Kanazawa. K-planes: Explicit radiance fields in space, time, and appearance. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12479–12488, 2023

    2023

  4. [12]

    C. Fu, G. Li, R. Song, W. Gao, and S. Liu. Octattention: Octree-based large-scale contexts model for point cloud compression. InProceedings of the AAAI conference on artificial intelligence, volume 36, pages 625–633, 2022

    2022

  5. [13]

    Geiger, P

    A. Geiger, P. Lenz, C. Stiller, and R. Urtasun. Vision meets robotics: The kitti dataset.The international journal of robotics research, 32(11):1231–1237, 2013

    2013

  6. [14]

    R. Gray. Vector quantization.IEEE Assp Magazine, 1(2):4–29, 1984

    1984

  7. [15]

    Hedman, J

    P. Hedman, J. Philip, T. Price, J.-M. Frahm, G. Drettakis, and G. Brostow. Deep blending for free-viewpoint image-based rendering.ACM Transactions on Graphics (ToG), 37(6):1–15, 2018

    2018

  8. [16]

    Kerbl, G

    B. Kerbl, G. Kopanas, T. Leimkühler, and G. Drettakis. 3d gaussian splatting for real-time radiance field rendering.ACM Trans. Graph., 42(4):139–1, 2023

    2023

  9. [17]

    Knapitsch, J

    A. Knapitsch, J. Park, Q.-Y . Zhou, and V . Koltun. Tanks and temples: Benchmarking large-scale scene reconstruction.ACM Transactions on Graphics (ToG), 36(4):1–13, 2017. 10

    2017

  10. [18]

    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, pages 21719–21728, 2024

    2024

  11. [19]

    L. Ling, Y . Sheng, Z. Tu, W. Zhao, C. Xin, K. Wan, L. Yu, Q. Guo, Z. Yu, Y . Lu, et al. Dl3dv-10k: A large-scale scene dataset for deep learning-based 3d vision. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 22160–22169, 2024

    2024

  12. [20]

    L. Liu, Z. Chen, and D. Xu. Hemgs: A hybrid entropy model for 3d gaussian splatting data compression. arXiv preprint arXiv:2411.18473, 2024

    2024 arXiv

  13. [21]

    X. Liu, X. Wu, S. Wang, Z. Li, and S. Kwong. Compgs++: Compressed gaussian splatting for static and dynamic scene representation.arXiv preprint arXiv:2504.13022, 2025

    2025 arXiv

  14. [22]

    X. Liu, X. Wu, P. Zhang, S. Wang, Z. Li, and S. Kwong. Compgs: Efficient 3d scene representation via compressed gaussian splatting. InProceedings of the 32nd ACM International Conference on Multimedia, pages 2936–2944, 2024

    2024

  15. [23]

    T. Lu, M. Yu, L. Xu, Y . Xiangli, L. Wang, D. Lin, and B. Dai. Scaffold-gs: Structured 3d gaussians for view-adaptive rendering. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 20654–20664, 2024

    2024

  16. [24]

    A. Luo, L. Song, K. Nonaka, K. Unno, H. Sun, M. Goto, and J. Katto. Scp: spherical-coordinate-based learned point cloud compression. InProceedings of the AAAI Conference on Artificial Intelligence, volume 38, pages 3954–3962, 2024

    2024

  17. [25]

    J. Ma, Y . Hu, L. Tang, J. Yang, Y . Zhai, and R. Wang. Enhancing 3d gaussian splatting compression via spatial condition-based prediction.arXiv preprint arXiv:2503.23337, 2025

    2025 arXiv

  18. [26]

    Mildenhall, P

    B. Mildenhall, P. P. Srinivasan, M. Tancik, J. T. Barron, R. Ramamoorthi, and R. Ng. Nerf: Representing scenes as neural radiance fields for view synthesis.Communications of the ACM, 65(1):99–106, 2021

    2021

  19. [27]

    Morgenstern, F

    W. Morgenstern, F. Barthel, A. Hilsmann, and P. Eisert. Compact 3d scene representation via self-organizing gaussian grids. InEuropean Conference on Computer Vision, pages 18–34. Springer, 2024

    2024

  20. [28]

    Navaneet, K

    K. Navaneet, K. Pourahmadi Meibodi, S. Abbasi Koohpayegani, and H. Pirsiavash. Compgs: Smaller and faster gaussian splatting with vector quantization. InEuropean Conference on Computer Vision, pages 330–349. Springer, 2024

    2024

  21. [29]

    Niedermayr, J

    S. Niedermayr, J. Stumpfegger, and R. Westermann. Compressed 3d gaussian splatting for accelerated novel view synthesis. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10349–10358, 2024

    2024

  22. [30]

    Papantonakis, G

    P. Papantonakis, G. Kopanas, B. Kerbl, A. Lanvin, and G. Drettakis. Reducing the memory footprint of 3d gaussian splatting.Proceedings of the ACM on Computer Graphics and Interactive Techniques, 7(1):1–17, 2024

    2024

  23. [31]

    R. Song, C. Fu, S. Liu, and G. Li. Efficient hierarchical entropy model for learned point cloud compression. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14368– 14377, 2023

    2023

  24. [32]

    H. Tang, S. Yang, Z. Liu, K. Hong, Z. Yu, X. Li, G. Dai, Y . Wang, and S. Han. Torchsparse++: Efficient training and inference framework for sparse convolution on gpus. InIEEE/ACM International Symposium on Microarchitecture (MICRO), 2023

    2023

  25. [33]

    Z. Tang, C. Feng, X. Cheng, W. Yu, J. Zhang, Y . Liu, X. Long, W. Wang, and L. Yuan. Neuralgs: Bridging neural fields and 3d gaussian splatting for compact 3d representations.arXiv preprint arXiv:2503.23162, 2025

    2025 arXiv

  26. [34]

    H. Wang, H. Zhu, T. He, R. Feng, J. Deng, J. Bian, and Z. Chen. End-to-end rate-distortion optimized 3d gaussian representation. InEuropean Conference on Computer Vision, pages 76–92. Springer, 2024

    2024

  27. [35]

    J. Wang, D. Ding, Z. Li, X. Feng, C. Cao, and Z. Ma. Sparse tensor-based multiscale representation for point cloud geometry compression.IEEE Transactions on Pattern Analysis and Machine Intelligence, 2022

    2022

  28. [36]

    Wang and W

    K. Wang and W. Gao. Unipcgc: Towards practical point cloud geometry compression via an efficient unified approach.Proceedings of the AAAI Conference on Artificial Intelligence, 39(12):12721–12729, Apr. 2025. 11

    2025

  29. [37]

    T. Wang, Z. Yu, and Y . Xu. Tc-gs: Tri-plane based compression for 3d gaussian splatting.arXiv preprint arXiv:2503.20221, 2025

    2025 arXiv

  30. [38]

    Y . Wang, Z. Li, L. Guo, W. Yang, A. Kot, and B. Wen. Contextgs: Compact 3d gaussian splatting with anchor level context model.Advances in neural information processing systems, 37:51532–51551, 2024

    2024

  31. [39]

    S. Xie, J. Liu, W. Zhang, S. Ge, S. Pan, C. Tang, Y . Bai, and Z. Wang. Sizegs: Size-aware compression of 3d gaussians with hierarchical mixed precision quantization.arXiv preprint arXiv:2412.05808, 2024

    2024

  32. [40]

    S. Xie, W. Zhang, C. Tang, Y . Bai, R. Lu, S. Ge, and Z. Wang. Mesongs: Post-training compression of 3d gaussians via efficient attribute transformation. InEuropean Conference on Computer Vision, pages 434–452. Springer, 2024

    2024

  33. [41]

    Z. Xu, X. Zhang, R. Li, Z. Tang, Q. Huang, and J. Zhang. Fakeshield: Explainable image forgery detection and localization via multi-modal large language models.arXiv preprint arXiv:2410.02761, 2024

    2024 arXiv

  34. [42]

    K. You, T. Chen, D. Ding, M. S. Asif, and Z. Ma. Reno: Real-time neural compression for 3d lidar point clouds.arXiv preprint arXiv:2503.12382, 2025

    2025 arXiv

  35. [43]

    Zhan, C.-Y

    Y .-T. Zhan, C.-Y . Ho, H. Yang, Y .-H. Chen, J. C. Chiang, Y .-L. Liu, and W.-H. Peng. Cat-3dgs: A context-adaptive triplane approach to rate-distortion-optimized 3dgs compression. InThe Thirteenth International Conference on Learning Representations

  36. [44]

    Zhan, H.-b

    Y .-T. Zhan, H.-b. Yang, C.-Y . Ho, J.-C. Chiang, and W.-H. Peng. Cat-3dgs pro: A new benchmark for efficient 3dgs compression.arXiv preprint arXiv:2503.12862, 2025

    2025 arXiv

  37. [45]

    Zhang and W

    C. Zhang and W. Gao. Adadpcc: Adaptive rate control and rate-distortion-complexity optimization for dynamic point cloud compression. InProceedings of the AAAI Conference on Artificial Intelligence, volume 39, pages 13188–13196, 2025

    2025

  38. [46]

    Gaussian

    X. Zhang, J. Meng, R. Li, Z. Xu, Y . Zhang, and J. Zhang. Gs-hider: Hiding messages into 3d gaussian splatting.arXiv preprint arXiv:2405.15118, 2024. 12 Technical Appendices A GausPcgc Framework A.1 Detailed Description We present the detailed architecture of our Four-stage Oc...

    2024 arXiv

Pith tools

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