Pith. sign in

REVIEW 4 major objections 6 minor 69 references

AtlasLC compresses object-centric 3D Gaussian splatting assets with no source data or retraining, cutting atlas preparation by up to 25x while preserving geometry at a small image-quality cost.

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

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

T0 review · deepseek-v4-flash

2026-08-01 13:58 UTC pith:OF4WRWHN

load-bearing objection AtlasLC is a solid, honest engineering paper on object-centric 3DGS compression whose speed/payload gains look real, but the opacity-only pruning demonstrably under-ranks thin/transparent silhouette regions, so the 'preserves object-wide foreground support' claim outruns the evidence. the 4 major comments →

arxiv 2607.26525 v1 pith:OF4WRWHN submitted 2026-07-29 cs.GR

AtlasLC: Fast Codec-Ready Compression of Object-Centric 3D Gaussian Splatting

ByungHyun Kim , Jinwoo Jeon , Woontack Woo This is my paper
classification cs.GR
keywords 3D Gaussian splattingobject-centric compressionatlas packinglocal-competition pruningextended realitydepth preservationsource-free compressiondeployment-aware evaluation
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

AtlasLC is a compression pipeline for 3D Gaussian Splatting (3DGS) assets of individual objects, aimed at the repeated packaging, transmission, and redeployment that XR object libraries require. The paper's claim is that object-centric 3DGS compression should be judged by a deployment-aware operating point—payload, preparation time, decode latency, runtime FPS, and preservation of silhouettes and depth—rather than bitrate alone. AtlasLC operates directly on a released Gaussian asset without original images, camera poses, or per-asset optimization, and it reports cutting atlas preparation by up to 25x and end-to-end compression time by up to 5x relative to prior atlas-based pipelines. On a 320-object evaluation set it achieves the lowest payload, lowest decode time, highest FPS, and highest 3D geometric fidelity among the compressed baselines, at a small PSNR cost (about 0.2-0.3 dB) and using roughly 6-8 percent fewer bits than similarly compact structured baselines. If correct, this makes scalable libraries of reusable radiance-field objects practical.

Core claim

AtlasLC's central discovery is that the mapping/remapping stage of structured 3DGS compression—not the bitrate—is the dominant bottleneck in object-centric XR asset preparation, and that this bottleneck can be removed by coupling two deterministic stages over one shared ordering. A lightweight single-pass sort-based conditional coordinate transform regularizes the unordered Gaussians into a stable 2D ordering; local-competition pruning then removes redundancy within coarse cells using opacity as the survivor score, and deterministic atlas packing places the retained splats into a codec-friendly atlas without any second map-generation stage. The reported result is a favorable deployment-aware

What carries the argument

The load-bearing object is the single-pass sort-based conditional coordinate backbone: a deterministic spherical parameterization followed by weighted CDF rank transforms (with opacity-based weights) that produce a stable 2D ordering of Gaussians. This ordering is quantized into coarse cells for local-competition pruning (each cell keeps a quota of the highest-opacity splats) and into packing bins for deterministic collision-free atlas placement (raster-scan slots, Morton-ordered points, global overflow spill). The same ordering is reused by both stages, eliminating any separate mapping/remapping step and bounding total cost by O(N log N) sorting.

Load-bearing premise

The load-bearing premise is that a splat's opacity alone is enough to decide which splats are locally redundant, so that keeping the highest-opacity splats in each cell preserves silhouettes, thin structures, and depth support; the paper's own failure cases show this can under-rank visually important low-opacity splats in thin, transparent, or high-anisotropy regions.

What would settle it

Compare AtlasLC against a version whose survivor score also accounts for scale or anisotropy on a set of objects with thin transparent structures (e.g., hair, lace, glass). If those assets show a significant drop in 3D F1 or spike in depth RMSE relative to the multi-attribute variant, the opacity-only priority is the limiting factor rather than packing.

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

Share X Bluesky LinkedIn Reddit HN

If this is right

  • Object assets can be compressed post-hoc, after release, without original capture data, so packaging and redeployment can be separated from the authoring pipeline.
  • Atlas preparation stops being a bottleneck: replacing iterative layout generation with a shared sort and deterministic packing brings map construction down to tens of milliseconds.
  • Geometry preservation improves when pruning is local rather than global: matched-budget comparisons show higher 3D F1 at aggressive pruning ratios, supporting silhouettes and depth cues.
  • The deployment advantage carries to a head-mounted-device setting: faster decode and higher FPS at the lowest payload among evaluated baselines, shortening estimated server-to-device transfer time.
  • The method transfers from virtual object datasets to real-world high-quality object captures, holding the lowest payload and total pipeline time with the highest FPS and 3D F1 among compressed baselines.

Where Pith is reading between the lines

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

  • A cheap multi-attribute survivor score (e.g., opacity weighted by scale or anisotropy) would likely recover the thin, transparent silhouette regions the paper identifies as failure cases, without sacrificing source-free operation; the paper's limitation section points to exactly this extension.
  • Because AtlasLC is source-free and training-free, its absolute rate-distortion is probably capped below what a source-assisted pipeline could reach; a small set of input views could close the PSNR gap while keeping the fast packaging.
  • The 'deployment-aware operating point' framing suggests new evaluation targets—startup latency, occlusion correctness, selection accuracy in live XR—that the paper does not test and that would determine whether the geometry gains translate into interaction benefits.
  • At-scale object libraries could exploit the preparation-time reduction to re-compress or re-package assets dynamically during streaming, a use case the paper mentions only as future work.

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

4 major / 6 minor

Summary. The paper proposes AtlasLC, a training-free, source-free compression pipeline for object-centric 3D Gaussian splatting. The method builds a single-pass sort-based conditional 2D ordering as a shared coordinate backbone, applies local-competition pruning with a fixed opacity-based survivor score, and packs the retained Gaussians into a deterministic, codec-friendly atlas that can be encoded with standard image/video codecs. The authors evaluate on 320 Objaverse objects and 30 Stanford-ORB objects, reporting lower payload, faster atlas preparation and end-to-end compression, higher rendering FPS, and higher 3D F1 than PLAS, LGSCV, UVGS, and FlexGaussian, at a small PSNR cost. The paper also includes a matched-count prune-only comparison against PUP, an ablation separating deterministic packing from local-competition pruning, an HMD-side deployment experiment, and an explicit limitations section.

Significance. If the results hold, AtlasLC addresses a real practical bottleneck: repeated, post-hoc packaging of reusable object-centric 3DGS assets for XR without source images or per-asset optimization. The paper's strengths are its clear problem framing, the deterministic and training-free design with O(N log N) complexity, the matched-budget prune-only comparison, the ablation isolating the two main components, and the honest limitations paragraph. The deployment-oriented evaluation with total pipeline time, decode latency, FPS, and geometry metrics is useful for engineering practice. However, the central geometry-preservation claim rests on a single opacity-based survivor score whose failure cases are admitted to include thin, transparent, and high-anisotropy silhouette-support regions; the global budget parameters are never reported or analyzed for sensitivity; and the depth-RMSE trends do not support an unconditional geometric advantage. These issues are load-bearing for the stated scope and need to be addressed before the claim as written is fully supported.

major comments (4)
  1. [§3.4.3 / §5] The load-bearing survivor score is s_i = α_i, and the same opacity-based weighting drives the shared ordering (§3.3.2). The paper's own §5 limitation admits that observed failure cases 'concentrate in thin, transparent, or high-anisotropy silhouette-support regions, where visually important low-opacity splats can be under-ranked.' These are precisely the regions where the paper's motivating geometry cues (silhouettes, thin structures, depth support) matter most for XR interaction. The aggregate 3D F1 gains over PLAS (+0.005, CI [+0.0014,+0.0089]) and LGSCV (+0.007, CI [+0.0036,+0.0107]) do not establish 'object-wide foreground support' for the object classes where the argument is most load-bearing. The authors should either provide stratified geometry metrics on thin/transparent/high-anisotropy assets or adopt a multi-attribute survivor score that incorporates scale or anisotropy while p
  2. [§3.4.2] The input-adaptive budget is defined by ρ_a = min(clip(M0/N_a, ρ_min, ρ_max), βT/N_a), but the numerical values of M0, ρ_min, ρ_max, β, and γ are never reported in the manuscript. The atlas resolution H×W, slot count K, competition-grid size, and packing-grid size are also fixed to a single configuration in all experiments. Because the rate, geometry, and timing comparisons in Tables 3–6 depend on this configuration, the claimed 6–8% payload advantage and the 'training-free' operating point are configuration-dependent. Please report the exact parameter values and include at least a one-dimensional sensitivity sweep (e.g., M0 or β versus payload/PSNR/3D F1) to show that the reported operating point is not a tuned artifact.
  3. [§4.2, Table 3] The geometry claim is supported by 3D F1, but the depth-RMSE metric in Table 3 is worse for AtlasLC (0.0361) than for PLAS (0.0350) and LGSCV (0.0354). The text states that depth-RMSE differences are trends whose bootstrap intervals cross zero, yet the abstract and conclusion claim a 'favorable deployment-aware balance ... of 3D geometry' and 'preserving object-wide foreground support.' Since depth fidelity is one of the paper's stated deployment cues, the geometry conclusion should be restricted to the 3D F1 measure, or the authors should provide evidence that depth behavior is not worse in the silhouette/thin-structure categories most relevant to XR. As it stands, the unconditional geometry-preservation language is stronger than the presented evidence.
  4. [Abstract / §4.2] The abstract's claim that AtlasLC 'uses about 6 to 8 percent fewer bits while maintaining comparable perceptual and geometric quality' is based on a single representative operating point (Table 3) where AtlasLC has lower PSNR than PLAS/LGSCV. The matched-payload interpolation in §4.2 equalizes payload, not quality; it does not demonstrate that AtlasLC achieves the same PSNR/LPIPS at lower bitrate. To substantiate the bitrate-at-comparable-quality claim, the authors should report bitrate at matched PSNR or matched LPIPS (or show that the rate-distortion curves cross in the appropriate direction). Otherwise the claim should be rephrased as 'lower payload at a small PSNR cost.'
minor comments (6)
  1. [§3.3.2] The notation 'bin(u_i)' in the conditional CDF v_i = F_{V|U}(v_i^0 | bin(u_i)) is not defined. Specify the coarse u-bin quantizer used for the conditional rank transform.
  2. [§3.4.2] M0 appears in the formula but is not explicitly named or defined before the equation. State clearly that M0 is a fixed keep-ratio numerator (possibly a target number of survivors), and give its units.
  3. [§4.1] There is a stray word 'material.' at the end of the experimental protocol paragraph; remove it.
  4. [Tables 3/4] The baseline labels are inconsistent: 'VANILLA' is typeset as 'V ANILLA' and 'FLEX*' is used in tables while the text uses 'FlexGaussian'. Standardize the notation.
  5. [§4.1 / Table 6] The 'GalaxyXR' HMD used for on-device evaluation is not described. Provide device specifications (SoC, RAM, renderer, GPU) so the HMD FPS/decode results are reproducible.
  6. [§5] The limitation section references a 'full failure taxonomy' in the supplementary material. Since the opacity-score failure is load-bearing, include at least a brief count or severity breakdown of the failure categories in the main text.

Circularity Check

0 steps flagged

AtlasLC is an empirical pipeline benchmark with no load-bearing prediction reducing to a fitted input or self-citation chain.

full rationale

The central claims are measured outcomes of a fixed, training-free pipeline. The survivor budget (§3.4.2) depends only on the input Gaussian count and fixed atlas capacity: ρa = min(clip(M0/Na, ρmin, ρmax), βT/Na), and the paper explicitly states that no PSNR, SSIM, LPIPS, 3D F1, depth, or per-object validation feedback is used. The survivor score (§3.4.3) is fixed as si = αi. Thus no output quantity is defined in terms of the metrics it is used to predict, and no fitted parameter is renamed as a prediction. The paper's own §5 limitation—that opacity-only selection under-ranks thin, transparent, or high-anisotropy silhouette-support splats—is an acknowledged correctness/generality risk, not a circular step. The only notable self-citation is [23] (OT-UVGS, by the first author), whose rank-based mapping resembles the §3.3.2 shared ordering backbone, but the paper does not invoke that citation to justify the backbone's correctness; the backbone is evaluated against independent baselines (PLAS, LGSCV, UVGS, FlexGaussian) and through ablations (Table 5). No uniqueness theorem is imported from the authors, no ansatz is smuggled in by citation, and the local-competition pruning plus deterministic packing combination is not merely a renamed existing result. Therefore circularity is minimal: a minor self-citation/attribution overlap, not load-bearing.

Axiom & Free-Parameter Ledger

7 free parameters · 5 axioms · 0 invented entities

AtlasLC introduces no new physical or generative entities: the atlas, competition cells, and packing bins are representations of existing Gaussian attributes. The load-bearing assumptions are the source-free sufficiency of the released asset, opacity as a pruning priority, codec-friendliness of atlas locality, and the local-redundancy model of object assets. The principal unaccounted free parameters are the global budget constants and grid sizes, whose reported values are absent.

free parameters (7)
  • M0 (keep-ratio numerator) = not reported
    Appears in ρa=min(clip(M0/Na,ρmin,ρmax), βT/Na) in §3.4.2 and directly controls the survivor count and therefore all downstream payload/quality results.
  • ρmin, ρmax (clip bounds) = not reported
    Global clip bounds on the keep ratio; no values or sensitivity analysis are given, so the operating point may be sensitive to their choice.
  • β (capacity multiplier) = not reported
    Used in the budget formula βT/Na; its choice trades payload against retained splats and is not reported.
  • γ (opacity weight exponent) = not reported
    The sort weights are wi=max(αi,ε)^γ in §3.3.2; γ is never given a value, although it shapes the shared ordering used by both pruning and packing.
  • Atlas resolution H×W and slots K = 512×512×1
    All experiments use a 512×512 atlas with K=1, fixing capacity T=262,144 and thus the maximum retained Gaussian count. No sensitivity to atlas size is reported.
  • Competition grid and packing grid sizes = 64×64 and 16×16
    The competition cell count and packing bin count are hand-chosen; they determine locality of pruning and overflow behavior, with no reported sensitivity.
  • Seam shift δ⋆ = per-asset search
    Selected to minimize weighted mass near the wrap boundary in §3.3.2. It is a deterministic per-asset adjustment rather than a quality-fitted parameter, but still an input-adaptive free choice.
axioms (5)
  • domain assumption A released 3DGS asset (Gaussian centers, opacities, attributes) is a complete representation for compression; original images, camera poses, and per-asset optimization are not required.
    The entire source-free setting in §1 and §3.1 depends on this. It is reasonable for post-hoc compression but is not derived.
  • domain assumption Opacity αi is a sufficient local survivor score for pruning.
    Used directly in §3.4.3 ('we use the fixed opacity score si=αi'). The paper's own §5 limitation section undercuts this for thin, transparent, and high-anisotropy regions.
  • domain assumption Atlas locality is codec-friendly; standard image/video codecs exploit local correlation in the packed atlas.
    The motivation for deterministic packing in §3.2 and §3.5 assumes that preserving spatial smoothness in the atlas yields compact HEVC encoding.
  • domain assumption Object-centric Gaussian redundancy is primarily local rather than globally disposable.
    This is the core justification for local-competition pruning in §2.3 and §3.4.1. It is plausible for object assets but not proven for all asset classes.
  • ad hoc to paper Largest-remainder apportionment with one survivor per non-empty cell preserves isolated support and distributes capacity across the object.
    The quota policy in §3.4.2 is a design heuristic; no formal argument is given that this allocation is optimal or that it always protects geometry-critical support.

pith-pipeline@v1.3.0-daily-deepseek · 19136 in / 13867 out tokens · 130404 ms · 2026-08-01T13:58:22.940072+00:00 · methodology

0 comments
Cite this review

Pith. "Pith review of AtlasLC: Fast Codec-Ready Compression of Object-Centric 3D Gaussian Splatting." pith.science (2026). https://pith.science/paper/OF4WRWHN

@misc{pith2026260726525,
  author       = {Pith},
  title        = {Pith review of: AtlasLC: Fast Codec-Ready Compression of Object-Centric 3D Gaussian Splatting},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/OF4WRWHN}},
  note         = {Machine review of arXiv:2607.26525}
}
Share X Bluesky LinkedIn Reddit HN
read the original abstract

3D Gaussian Splatting (3DGS) enables photorealistic novel-view synthesis with real-time rendering, but deploying compressed object-centric 3DGS in XR requires more than image-space rate-distortion. In practical XR asset pipelines, reusable objects are repeatedly packaged, transmitted, decoded, and instantiated, making asset-preparation cost, codec compatibility, decoding latency, and preservation of depth and silhouette cues first-class concerns. Existing 3DGS compression methods are largely developed for scene-scale captures and often rely on heavy layout generation or aggressive global pruning, assumptions that transfer poorly to semantically concentrated foreground objects. We present AtlasLC, a source-free, training-free compression pipeline for object-centric 3DGS that operates directly on released Gaussian assets, without original images, camera poses, or per-asset optimization. AtlasLC couples local-competition pruning with deterministic atlas packing to remove the mapping/remapping bottleneck while preserving object-wide foreground support; a lightweight single-pass sort-based conditional transport is used as a shared coordinate backbone for these stages. Across the evaluated assets, AtlasLC reduces atlas-preparation time by up to a factor of 25 and end-to-end compression time by up to a factor of 5, while offering a favorable deployment-aware balance of payload, decode latency, runtime FPS, and 3D geometry relative to the evaluated compressed baselines. Relative to similarly compact structured baselines, it uses about 6 to 8 percent fewer bits while maintaining comparable perceptual and geometric quality. These results show that object-centric 3DGS compression should be optimized for a deployment-aware operating point enabling scalable XR asset libraries.

Figures

Figures reproduced from arXiv: 2607.26525 by Byunghyun Kim, Jinwoo Jeon, Woontack Woo.

Figure 1
Figure 1. Figure 1: Pipeline of AtlasLC for object-centric 3DGS compression. AtlasLC first builds a single-pass sort-based conditional coordinate system over the input Gaussians, then applies local-competition pruning within coarse coordinate cells to remove locally redundant splats while preserving object-wide foreground support (a – d). The retained Gaussians are finally deterministically packed into a codec-friendly atlas … view at source ↗
Figure 2
Figure 2. Figure 2: Rate-dependent fidelity, geometry components, and de￾ployment cost. We sweep HEVC QP {1,4,8,12,16,20} for AtlasLC, PLAS, LGSCV, and UVGS, and use the native rate ladder for Flex￾Gaussian. Payload–PSNR shows the image-space fidelity trade-off, 3D precision and recall decompose geometry preservation, and payload– total time summarizes end-to-end deployment cost. Within the common AtlasLC/PLAS/LGSCV payload r… view at source ↗
Figure 3
Figure 3. Figure 3: Visualization of prune-only behavior on an object-centric 3DGS asset. [PITH_FULL_IMAGE:figures/full_fig_p005_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: Representative RGB comparison on three Objaverse assets. We compare VANILLA, PLAS, AtlasLC (Ours), UVGS, and FLEX using RGB renderings. AtlasLC remains visually close to VANILLA and PLAS, while UVGS and FLEX* show slight boundary degradation in thin structures (e.g., the narrow ankle region of the boot). Corresponding depth-error visualizations are provided in [PITH_FULL_IMAGE:figures/full_fig_p007_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: Representative depth-error comparison on three Objaverse assets. We compare VANILLA, PLAS, AtlasLC (Ours), UVGS, and FLEX using per-pixel depth-error visualizations. AtlasLC exhibits cleaner depth boundaries and lower depth error than UVGS and FLEX, particularly around thin and detailed structures. Methods are shown at the representative operating points used in [PITH_FULL_IMAGE:figures/full_fig_p008_5.png] view at source ↗

discussion (0)

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

Reference graph

Works this paper leans on

69 extracted references · 10 canonical work pages · 1 internal anchor

  1. [1]

    Traffic Characteristics of Extended Reality

    A. Alnajim, S. Salehi, C.-C. Shen, and M. Smith. Traffic characteristics of extended reality.arXiv preprint arXiv:2304.07908, 2023. doi: 10. 48550/arXiv.2304.079082, 3

    work page internal anchor Pith review Pith/arXiv arXiv doi:10.48550/arxiv.2304.07908 2023

  2. [2]

    M. T. Bagdasarian, P. Knoll, Y .-H. Li, F. Barthel, A. Hilsmann, P. Eisert et al. 3DGS.zip: A survey on 3d gaussian splatting compression methods. Computer Graphics Forum, 2025. doi:10.1111/cgf.700782, 3

    work page doi:10.1111/cgf.700782 2025

  3. [3]

    J. Cao, V . Goel, C. Wang, A. Kag, J. Hu, S. Korolev et al. Lightweight predictive 3d gaussian splats. InThe Thirteenth International Con- ference on Learning Representations (ICLR), 2025. Also available as arXiv:2406.19434. 2, 3

    Pith/arXiv arXiv 2025

  4. [4]

    Y . Chen, Q. Wu, M. Li, W. Lin, M. Harandi, and J. Cai. Fast feedfor- ward 3d gaussian splatting compression. InThe Thirteenth International Conference on Learning Representations (ICLR), 2025. Also available as arXiv:2410.08017. 2, 3

    Pith/arXiv arXiv 2025

  5. [5]

    Y . Chen, Q. Wu, W. Lin, M. Harandi, and J. Cai. Hac: Hash-grid assisted context for 3d gaussian splatting compression. InComputer Vision – ECCV 2024, pp. 422–438. Springer, 2024. doi: 10.1007/978-3-031-72667-5_24 2, 3

    work page doi:10.1007/978-3-031-72667-5_24 2024

  6. [6]

    Deitke, D

    M. Deitke, D. Schwenk, J. Salvador, L. Weihs, O. Michel, E. VanderBilt et al. Objaverse: A universe of annotated 3d objects. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pp. 13142–13153, 2023. 2, 3, 6

    2023

  7. [7]

    Di Sario, R

    F. Di Sario, R. Renzulli, M. Grangetto, A. Sugimoto, and E. Tartaglione. Gode: Gaussians on demand for progressive level of detail and scalable compression.arXiv preprint arXiv:2501.13558, 2025. 3

    arXiv 2025

  8. [8]

    M. D. Dogan, E. J. Gonzalez, K. Ahuja, R. Du, A. Colaço, J. Lee et al. Augmented object intelligence with xr-objects. InProceedings of the 37th Annual ACM Symposium on User Interface Software and Technology (UIST ’24), pp. 1–15. Association for Computing Machinery, 2024. Also available as arXiv:2404.13274. doi:10.1145/3654777.36763792, 3

    Pith/arXiv arXiv 2024

  9. [9]

    Z. Dong, K. Chen, Z. Lv, H.-X. Yu, Y . Zhang, C. Zhang et al. Digital twin catalog: A large-scale photorealistic 3d object digital twin dataset. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pp. 753–763, June 2025. 2, 3

    2025

  10. [10]

    X. Du, Y . Wang, K. Zhan, and X. Yu. Mobile-gs: Real-time gaussian splatting for mobile devices. InThe Fourteenth International Conference on Learning Representations (ICLR), 2026. ICLR 2026 Poster. 1, 2

    2026

  11. [11]

    Z. Fan, K. Wang, K. Wen, Z. Zhu, D. Xu, and Z. Wang. Lightgaussian: Unbounded 3d gaussian compression with 15x reduction and 200+ fps. In Advances in Neural Information Processing Systems 37 (NeurIPS), 2024. Also available as arXiv:2311.17245. 2, 3

    Pith/arXiv arXiv 2024

  12. [12]

    Franke, L

    L. Franke, L. Fink, and M. Stamminger. Vr-splatting: Foveated radiance field rendering via 3d gaussian splatting and neural points.Proceedings of the ACM on Computer Graphics and Interactive Techniques, 8(1), art. no. 18, May 2025. doi:10.1145/37283021, 2

    work page doi:10.1145/37283021 2025

  13. [13]

    Friston, C

    S. Friston, C. Fan, J. Doboš, T. Scully, and A. Steed. 3drepo4unity: Dynamic loading of version controlled 3d assets into the unity game engine. InProceedings of the 22nd International Conference on 3D Web Technology (Web3D ’17), art. no. 15, 9 pp. Association for Computing Machinery, 2017. doi:10.1145/3055624.30759412, 3

    arXiv 2017

  14. [14]

    S. Gao, J. Liu, Q. Jiang, F. Sinclair, W. Sentosa, B. Godfrey et al. Xrgo: Design and evaluation of rendering offload for low-power extended reality devices. InProceedings of the 16th ACM Multimedia Systems Conference (MMSys ’25), pp. 124–135, 2025. doi:10.1145/3712676.37144442, 3

    arXiv 2025

  15. [15]

    Guédon and V

    A. Guédon and V . Lepetit. Gaussian frosting: Editable complex radiance fields with real-time rendering. InComputer Vision – ECCV 2024, pp. 413–

    2024

  16. [16]

    Guédon and V

    A. Guédon and V . Lepetit. Sugar: Surface-aligned gaussian splatting for efficient 3d mesh reconstruction and high-quality mesh rendering. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pp. 5354–5363, June 2024. 2, 3

    2024

  17. [17]

    J. Guo, Y . Xin, G. Liu, K. Xu, L. Liu, and R. Hu. Articulatedgs: Self- supervised digital twin modeling of articulated objects using 3d gaussian splatting. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pp. 27144–27153, June 2025. 2

    2025

  18. [18]

    Hanson, A

    A. Hanson, A. Tu, V . Singla, M. Jayawardhana, M. Zwicker, and T. Gold- stein. Pup 3d-gs: Principled uncertainty pruning for 3d gaussian splatting. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pp. 5949–5958, June 2025. Also available as arXiv:2406.10219. 2, 3, 4, 5

    arXiv 2025

  19. [19]

    Hellmuth, M

    C.-P. Hellmuth, M. Bachinski, and J. Müller. Interaction techniques for 3d-positioning objects in mobile augmented reality. InICMI ’21: International Conference on Multimodal Interaction, pp. 604–612. ACM,

  20. [20]

    Huang, Z

    B. Huang, Z. Yu, A. Chen, A. Geiger, and S. Gao. 2d gaussian splatting for geometrically accurate radiance fields. InSIGGRAPH 2024 Conference Papers, art. no. 32, 11 pp. Association for Computing Machinery, 2024. doi:10.1145/3641519.36574282, 3

    arXiv 2024

  21. [21]

    Kerbl, G

    B. Kerbl, G. Kopanas, T. Leimkühler, and G. Drettakis. 3d gaussian splatting for real-time radiance field rendering.ACM Transactions on Graphics, 42(4):139:1–139:14, 2023. doi:10.1145/35924331, 2, 6

    doi:10.1145/35924331 2023

  22. [22]

    B. Kim. Semantic-Fast-SAM: Efficient semantic segmenter. In2025 Asia Pacific Signal and Information Processing Association Annual Summit and Conference (APSIPA ASC), pp. 1158–1163. IEEE, 2025. doi: 10. 1109/APSIPAASC65261.2025.112493159

    arXiv 2025

  23. [23]

    B. Kim. OT-UVGS: Revisiting UV mapping for gaussian splatting as a capacity allocation problem. InEurographics 2026 – Short Papers. The Eurographics Association, 2026. doi:10.2312/egs.202610233

    work page doi:10.2312/egs.202610233 2026

  24. [24]

    Kim, B.-Y

    D.-H. Kim, B.-Y . Choi, K.-J. Oh, G. Lee, and J.-G. Kim. A vq-hevc hybrid compression method for learned 3dgs (3d gaussian splatting) models. Journal of Broadcast Engineering, 30(7):1155–1166, December 2025. doi: 10.5909/JBE.2025.30.7.11552, 3

    work page doi:10.5909/jbe.2025.30.7.11552 2025

  25. [25]

    S.-Y . Kim, D. Kim, W. Cho, H. Song, S. Kang, and W. Woo. Scenelinker: Compositional 3d scene generation via semantic scene graph from rgb sequences.IEEE Transactions on Visualization and Computer Graphics, 32(5):3125–3135, 2026. doi:10.1109/TVCG.2026.36798933

    arXiv 2026

  26. [26]

    Kuang, Y

    Z. Kuang, Y . Zhang, H.-X. Yu, S. Agarwala, S. Wu, and J. Wu. Stanford- ORB: A real-world 3d object inverse rendering benchmark. InAdvances in Neural Information Processing Systems, vol. 36, pp. 46938–46957. Curran Associates, Inc., December 2023. NeurIPS 2023 Datasets and Benchmarks Track. 6

    2023

  27. [28]

    S. Lee, F. Shu, Y . Sanchez, T. Schierl, and C. Hellge. Compression of 3d gaussian splatting with optimized feature planes and standard video codecs. InProceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pp. 25496–25505, October 2025. 2, 3

    2025

  28. [29]

    K. Li, M. Masuda, S. Schmidt, and S. Mori. Radiance fields in XR: A survey on how radiance fields are envisioned and addressed for XR research.IEEE Transactions on Visualization and Computer Graphics, 31(11):9709–9719, 2025. doi:10.1109/TVCG.2025.36167941, 2, 3

    arXiv 2025

  29. [30]

    Li, Y .-S

    S. Li, Y .-S. Liu, and Z. Han. Gaussianudf: Inferring unsigned distance functions through 3d gaussian splatting. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pp. 27113–27123, June 2025. 2, 3

    2025

  30. [31]

    Lilija, H

    K. Lilija, H. Pohl, S. Boring, and K. Hornbæk. Augmented reality views for occluded interaction. InProceedings of the 2019 CHI Conference on Human Factors in Computing Systems, pp. 1–12. Association for Comput- ing Machinery, 2019. doi:10.1145/3290605.33006762, 3

    arXiv 2019

  31. [32]

    X. Liu, P. Tayal, J. Wang, J. Zarzar, T. Monnier, K. Tertikas et al. Un- common objects in 3d. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pp. 14102–14113, June

  32. [33]

    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, pp. 2936–

  33. [34]

    Y . Liu, B. Han, F. Qian, A. Narayanan, and Z.-L. Zhang. Vues: Prac- tical mobile volumetric video streaming through multiview transcoding. InProceedings of the 28th Annual International Conference on Mobile Computing and Networking (MobiCom ’22), pp. 514–527, 2022. doi: 10. 1145/3495243.35170272, 3

    arXiv 2022

  34. [35]

    Y . Liu, B. Jia, R. Lu, J. Ni, S.-C. Zhu, and S. Huang. Building interactable replicas of complex articulated objects via gaussian splatting. InThe Thirteenth International Conference on Learning Representations (ICLR),

  35. [36]

    Q. Luo, H. Zhang, M. Xu, B. Di, A. Chen, S. Mao et al. An overview of 3gpp standardization for extended reality (xr) in 5g and beyond.GetMo- bile: Mobile Computing and Communications, 27(3):10–17, 2023. doi: 10. 1145/3631588.36315922, 3

    arXiv 2023

  36. [37]

    S. Ma, Y . Luo, W. Yang, and Y . Yang. Mags: Reconstructing and sim- 10 To appear in IEEE Transactions on Visualization and Computer Graphics. ulating dynamic 3d objects with mesh-adsorbed gaussian splatting. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pp. 8745–8755, October 2025. 2, 3

    2025

  37. [38]

    M. C. F. Macedo and A. L. Apolinario. Occlusion handling in augmented reality: Past, present, and future.IEEE Transactions on Visualization and Computer Graphics, 29(2):1590–1609, 2023. doi: 10.1109/TVCG.2021. 31178662, 3

    doi:10.1109/tvcg.2021 2023

  38. [39]

    Morgenstern, F

    W. Morgenstern, F. Barthel, A. Hilsmann, and P. Eisert. Compact 3d scene representation via self-organizing gaussian grids. InComputer Vision – ECCV 2024, pp. 18–34. Springer, 2024. Also available as arXiv:2312.13299. doi:10.1007/978-3-031-73013-9_22, 3, 6

    Pith/arXiv arXiv 2024

  39. [40]

    K. L. Navaneet, K. Pourahmadi Meibodi, S. Abbasi Koohpayegani, and H. Pirsiavash. Compgs: Smaller and faster gaussian splatting with vector quantization. InComputer Vision – ECCV 2024, pp. 330–349. Springer,

    2024

  40. [41]

    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 (CVPR), pp. 10349–10358, June 2024. doi: 10.1109/CVPR52733.2024.00985 2, 3

    arXiv 2024

  41. [42]

    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), art. no. 16, 2024. Also available as arXiv:2406.17074. doi:10.1145/36512822, 3

    Pith/arXiv arXiv 2024

  42. [43]

    A. Rai, D. Wang, M. Jain, N. Sarafianos, K. Chen, S. Sridhar et al. Uvgs: Reimagining unstructured 3d gaussian splatting using uv mapping. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pp. 5927–5937, June 2025. 2, 3, 6

    2025

  43. [44]

    Y . Shi, G. Morin, S. Gasparini, and W. T. Ooi. Lapisgs: Layered progres- sive 3d gaussian splatting for adaptive streaming. InProceedings of the 2025 International Conference on 3D Vision (3DV), pp. 991–1000. IEEE,

    2025

  44. [45]

    J. Tang, Z. Chen, X. Chen, T. Wang, G. Zeng, and Z. Liu. LGM: Large multi-view gaussian model for high-resolution 3D content creation. In Computer Vision – ECCV 2024, vol. 15062 ofLecture Notes in Computer Science, pp. 1–18. Springer, Cham, 2025. doi:10.1007/978-3-031-73235-5_1 9

    work page doi:10.1007/978-3-031-73235-5_1 2024

  45. [46]

    X. Tang, X. Hu, C.-W. Fu, and D. Cohen-Or. Grabar: Occlusion-aware grabbing virtual objects in ar. InProceedings of the 33rd Annual ACM Symposium on User Interface Software and Technology, pp. 697–708. As- sociation for Computing Machinery, 2020. doi: 10.1145/3379337.3415835 2, 3

    arXiv 2020

  46. [47]

    B. Tian, Q. Gao, S. Xianyu, X. Cui, and M. Zhang. Flexgaussian: Flexible and cost-effective training-free compression for 3d gaussian splatting. In Proceedings of the 33rd ACM International Conference on Multimedia, pp. 7287–7296. Association for Computing Machinery, 2025. Also available as arXiv:2507.06671. doi:10.1145/3746027.37547442, 3, 6, 8

    Pith/arXiv arXiv 2025

  47. [48]

    Y .-Z. Tsai, X. Zhang, Z. Li, and J. Chen. L3gs: Layered 3d gaussian splats for efficient 3d scene delivery. InProceedings of the 31st Annual Interna- tional Conference on Mobile Computing and Networking. Association for Computing Machinery, 2025. Also available as arXiv:2504.05517. doi: 10 .1145/3680207.37234852, 3

    Pith/arXiv arXiv 2025

  48. [49]

    D. R. Walton and A. Steed. Accurate real-time occlusion for mixed reality. InProceedings of the 23rd ACM Symposium on Virtual Reality Software and Technology, pp. 1–10. Association for Computing Machinery, 2017. doi:10.1145/3139131.31391532, 3

    arXiv 2017

  49. [50]

    doi: 10.1109/3DV66043.2025

    Also available as arXiv:2408.14823. doi: 10.1109/3DV66043.2025. 000962, 3

    arXiv 2025

  50. [51]

    T. Wang, M. Li, G. Zeng, C. Meng, and Q. Zhang. Gaussian herding across pens: An optimal transport perspective on global gaussian reduction for 3dgs. InAdvances in Neural Information Processing Systems (NeurIPS),

  51. [52]

    Y . Wang, Z. Li, L. Guo, W. Yang, A. C. Kot, and B. Wen. Contextgs: Compact 3d gaussian splatting with anchor level context model. InAd- vances in Neural Information Processing Systems 37 (NeurIPS), 2024. Also available as arXiv:2405.20721. 2, 3

    Pith/arXiv arXiv 2024

  52. [53]

    S. Xie, J. Liu, W. Zhang, S. Ge, S. Pan, C. Tang et al. Sizegs: Size-aware compression of 3d gaussian splatting via mixed integer programming. In Proceedings of the 33rd ACM International Conference on Multimedia. Association for Computing Machinery, 2025. Best Paper Candidate. doi: 10.1145/3746027.37553703

    arXiv 2025

  53. [54]

    S. Xie, W. Zhang, C. Tang, Y . Bai, R. Lu, S. Ge et al. Mesongs: Post- training compression of 3d gaussians via efficient attribute transformation. InComputer Vision – ECCV 2024, pp. 434–452. Springer, 2024. Also available as arXiv:2409.09756. doi:10.1007/978-3-031-73414-4_252, 3

    Pith/arXiv arXiv 2024

  54. [55]

    Q. Yang, G. Van der Auwera, and Z. Li. Lightweight 3d gaussian splatting compression via video codec.arXiv preprint arXiv:2512.11186, 2025. arXiv preprint. 2, 3, 6

    arXiv 2025

  55. [56]

    H. Wang, H. Zhu, T. He, R. Feng, J. Deng, J. Bian et al. End-to-end rate-distortion optimized 3d gaussian representation. InComputer Vision – ECCV 2024, pp. 76–92. Springer, 2024. doi: 10.1007/978-3-031-73636-0_5 3

    work page doi:10.1007/978-3-031-73636-0_5 2024

  56. [57]

    Z. Yang, J. Wang, H. Zhang, S. Manivasagam, Y . Chen, and R. Urtasun. Genassets: Generating in-the-wild 3d assets in latent space. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pp. 22392–22403, June 2025. 2, 3

    2025

  57. [58]

    NeurIPS 2025 Spotlight. 3

    2025

  58. [59]

    Zhang, C

    B. Zhang, C. Fang, R. Shrestha, Y . Liang, X. Long, and P. Tan. Rade-gs: Rasterizing depth in gaussian splatting.ACM Transactions on Graphics,

  59. [60]

    Zhang, T

    Z. Zhang, T. Song, Y . Lee, L. Yang, C. Peng, R. Chellappa et al. Lp- 3dgs: Learning to prune 3d gaussian splatting. InAdvances in Neural Information Processing Systems 37 (NeurIPS), 2024. Also available as arXiv:2405.18784. 2, 3

    Pith/arXiv arXiv 2024

  60. [61]

    X. Zhao, W. Ding, Y . An, Y . Du, T. Yu, M. Li et al. Fast segment anything,

  61. [63]

    Q. Yang, K. Yang, Y . Xing, Y . Xu, and Z. Li. A benchmark for gaussian splatting compression and quality assessment study. InProceedings of the 6th ACM International Conference on Multimedia in Asia, MMAsia ’24, art. no. 12, 8 pp. Association for Computing Machinery, 2024. doi: 10 .1145/3696409.37001722, 3

    arXiv 2024

  62. [65]

    Z. Ye, C. Wan, C. Li, J. Hong, S. Li, L. Li et al. 3d gaussian rendering can be sparser: Efficient rendering via learned fragment pruning. InAdvances in Neural Information Processing Systems 37 (NeurIPS), 2024. 3

    2024

  63. [430]

    ECCV 2024 Oral

    Springer, 2024. ECCV 2024 Oral. doi: 10.1007/978-3-031-73650-6_24 2, 3

    work page doi:10.1007/978-3-031-73650-6_24 2024

  64. [2021]

    doi:10.1145/3462244.34799452, 3

    arXiv

  65. [2023]

    doi:10.48550/arXiv.2306.121569 11

    Pith/arXiv arXiv doi:10.48550/arxiv.2306.12156

  66. [2024]

    doi:10.1007/978-3-031-73411-3_192, 3

    work page doi:10.1007/978-3-031-73411-3_192

  67. [2025]

    Published as a conference paper at ICLR 2025. 2, 3

    2025

  68. [2026]

    doi: 10.1145/3789201 2, 6

    Online publication date: January 16, 2026. doi: 10.1145/3789201 2, 6

    doi:10.1145/3789201 2026

  69. [2944]

    doi: 10.1145/3664647

    Association for Computing Machinery, 2024. doi: 10.1145/3664647. 36814683

    doi:10.1145/3664647 2024