Pith. sign in

REVIEW 2 major objections 5 minor 67 references

Open-vocabulary 3D Gaussian fields can be built without training and stored under a megabyte by coding only the binding to a small set of spatial anchors.

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 · grok-4.5

2026-07-14 13:18 UTC pith:IB47SEAW

load-bearing objection Training-free sub-MB open-vocab 3DGS fields that actually match trained baselines by coding the binding, not the features. the 2 major comments →

arxiv 2607.10237 v1 pith:IB47SEAW submitted 2026-07-11 cs.CV

CoSAG: Compact Semantic Anchor Gaussians via Training-Free Rate-Distortion Coding

Yuang Jia , Jinlong Wang , Junhong Lin , Ruiting Dai , Wei Gao This is my paper
classification cs.CV
keywords 3D Gaussian Splattingopen-vocabulary segmentationsemantic field compressionrate-distortion codingtraining-free constructionCLIP featuresanchor binding
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.

Attaching a high-dimensional vision-language feature to every Gaussian in a 3D scene balloons storage into the gigabytes and blocks practical use. Prior compact methods still train a per-scene codec or codebook and leave the per-Gaussian assignment uncompressed. This paper separates construction from storage: it lifts features onto a frozen Gaussian scene with a closed-form transmittance-weighted average, binds each Gaussian to a small table of spatially grounded semantic anchors, cleans the anchors by multi-view averaging, and then entropy-codes the resulting assignment along a space-filling curve. Because nearby Gaussians share an anchor, the binding is highly predictable, so almost all of the rate goes to residual errors rather than noise. The result is a single representation that serves 2D rendering, direct 3D selection, and scene editing while staying under a megabyte and matching or beating previous accuracy.

Core claim

Storage of an open-vocabulary semantic Gaussian field is a rate-distortion problem over the per-Gaussian binding to a small, spatially grounded anchor table. Once construction is performed by a closed-form lift and multi-view denoising, that binding becomes clean and piecewise constant in space, so a predictive entropy coder (with parent-pointer sharing across hierarchy levels) drives the entire field into the sub-megabyte regime without shipping any decoder, while accuracy equals or exceeds optimization-based methods.

What carries the argument

Spatially grounded semantic anchors plus rate-distortion binding compression: each Gaussian is bound by joint spatial-semantic proximity to an anchor obtained from multi-view SAM regions; the binding is then Morton-reordered and entropy-coded, with coarser hierarchy levels recovered by parent-pointer tables.

Load-bearing premise

The method assumes that depth rendered from the frozen Gaussian scene is reliable enough to place the anchors in 3D; if the geometry is wrong, both the spatial binding and its compressibility collapse.

What would settle it

On a scene whose reconstructed depth is known to be unreliable (for example, large transparent or reflective surfaces), measure whether CoSAG's mIoU and storage both degrade relative to a method that does not back-project through that depth; a clear drop would falsify the grounding claim.

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

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

2 major / 5 minor

Summary. CoSAG constructs open-vocabulary semantic fields on frozen 3D Gaussian Splatting by a closed-form transmittance-weighted lift of multi-granularity SAM-CLIP features, binds Gaussians to spatially grounded anchors via joint spatial-semantic proximity, denoises anchors by multi-view averaging with contrast-based level selection, and stores the result by deterministic shared-basis projection of the anchor table plus Morton-ordered, parent-pointer entropy coding of the binding. Construction and storage are fully decoupled and training-free; no per-scene decoder is shipped. On LERF-OVS, 3D-OVS, LERF 3D-selection, and Replica LSeg, the method reports sub-megabyte fields that match or exceed recent autoencoder, codebook, and distilled-field baselines (including LangSplatV2, SCOUP, CF3, LightSplat) while reducing storage by roughly 37–76× relative to LangSplatV2, with supporting ablations of binding, multi-view averaging, coder stages, and projection dimension.

Significance. If the reported numbers hold, the paper cleanly separates field construction from field storage and shows that the per-Gaussian binding—not the high-dimensional features themselves—is the compressible object. The closed-form lift, spatially grounded anchors, and decoder-free entropy coder together yield a practical sub-MB representation that serves both 2D-rendered and 3D-selection protocols from one field, plus scene editing. Strengths include multi-protocol evaluation against strong recent baselines, fixed hyperparameters across scenes, ablations that isolate binding and coding contributions (Tables 6–7, Fig. 4), and a transfer experiment from SAM-CLIP to dense LSeg. The work is a concrete systems advance for deployable open-vocabulary 3DGS rather than a purely incremental accuracy paper.

major comments (2)
  1. Sec. 3.3 and Conclusion: anchors are grounded by back-projecting SAM-region pixels through depth rendered from the frozen 3DGS. The authors correctly flag this as a limitation where geometry is unreliable. The main tables (LERF-OVS, 3D-OVS, Replica) use indoor scenes with reliable pretrained 3DGS; the appendix Waymo result is qualitative only. A short quantitative stress test (e.g., controlled depth noise or a scene with known reconstruction failure) would strengthen the claim that the reported compressibility and accuracy transfer beyond the evaluated indoor setting. This does not overturn the indoor results but is load-bearing for the generality of the rate–distortion argument.
  2. Sec. 3.3, Eq. (3): the method minimizes a decoupled transmittance-weighted surrogate rather than the coupled rendering objective whose Gram matrix couples all Gaussians. The paper presents this as admitting a unique closed-form minimizer and reports strong empirical accuracy, but does not quantify the approximation gap (e.g., residual of the true rendering loss after the closed-form lift, or a controlled comparison against a few steps of the coupled objective). A brief residual or side-by-side check would make the sufficiency claim more rigorous; the current multi-dataset tables support practicality but leave the theoretical gap unmeasured.
minor comments (5)
  1. Sec. 4.1 vs. Sec. 3.3: the cosine gate is written as θ in the method and as τ=0.7 in implementation details; unify the symbol.
  2. Table 1 overall mAcc: CoSAG is second to SCOUP (86.1 vs 86.9) while leading mIoU; the abstract’s “matching or exceeding” is accurate but a one-sentence note in the text would avoid over-reading the mAcc column.
  3. Fig. 2 pipeline diagram is dense; a short caption call-out of which stages are closed-form vs. which are entropy-coded would help readers unfamiliar with the LangSplat lineage.
  4. Related work: concurrent LightSplat and SCOUP are discussed; ensure arXiv dates and protocol differences (single-level vs multi-granularity, uncompressed index vs coded binding) are stated clearly so the novelty claim on binding compression remains unambiguous.
  5. Typographical: occasional missing spaces after commas and in compound adjectives (e.g., “rate–distortion” vs “rate-distortion”) appear in the abstract and introduction; a light copy-edit pass would clean them.

Circularity Check

0 steps flagged

No circularity: empirical systems method with closed-form construction and external benchmark metrics; nothing reduces to its own inputs by definition.

full rationale

CoSAG’s load-bearing chain is (i) a closed-form transmittance-weighted lift (Eq. 3) of frozen 3DGS compositing weights against off-the-shelf SAM/CLIP (or LSeg) maps, (ii) learning-free spatial–semantic binding to anchors, (iii) multi-view scatter-mean denoising, and (iv) deterministic PCA plus Morton/LZMA/parent-pointer coding of the binding. None of these steps defines the reported quantities in terms of themselves: mIoU, mAcc, mAcc@0.25, and storage MB are measured against public LERF-OVS, 3D-OVS, and Replica annotations and compared to independent baselines. Hyperparameters (θ=0.7, κ=2000, ρ=10^{-4}, D=128/32/16) are fixed a priori across scenes rather than fitted to the reported scores and then re-presented as predictions. There is no uniqueness theorem, self-citation chain, or ansatz smuggled in as a first-principles derivation; self-citations (if any) are not load-bearing for the central claim. The depth-grounding premise is an acknowledged limitation, not a circular reduction. Score 0 is therefore appropriate.

Axiom & Free-Parameter Ledger

5 free parameters · 4 axioms · 2 invented entities

Load-bearing content is a CV systems method. Free parameters are the fixed gates and dimensions chosen once for all scenes. Axioms are standard domain assumptions (frozen 3DGS, SAM/CLIP teachers, alpha compositing). Invented entities are algorithmic constructs (semantic anchors, parent-pointer hierarchy, Morton binding stream), not physical postulates; they have no independent evidence outside the reported benchmarks.

free parameters (5)
  • cosine gate θ = 0.7
    Semantic agreement threshold for candidate anchors; fixed at 0.7 for all scenes (Sec. 4.1).
  • neighborhood size κ = 2000
    Number of nearest sampling points considered for binding; fixed at 2000 (Sec. 4.1).
  • singleton promotion fraction ρ = 1e-4
    Fraction of high-variance Gaussians promoted to singleton anchors; fixed at 1e-4 (Sec. 3.3).
  • per-level projection dimensions D = 128/32/16
    Shared-basis ranks for whole/part/subpart levels; chosen as 128/32/16 by semantic richness (Sec. 3.5, Fig. 4).
  • relevancy box size and consistency exponent γ = 29×29, γ=2
    29×29 box average and γ=2 used in contrast-based level selection and related consistency (Sec. 3.4, 4.1).
axioms (4)
  • domain assumption A frozen pretrained 3DGS provides accurate enough geometry and transmittance weights for closed-form feature lift and depth-based anchor grounding.
    Invoked throughout Sec. 3.1–3.3; construction never updates Gaussians.
  • domain assumption Off-the-shelf SAM multi-granularity masks plus CLIP (or LSeg) region/pixel features are adequate teachers for open-vocabulary 3D semantics.
    Feature extraction pipeline in Sec. 3.1; dense LSeg variant in Sec. 4.4.
  • domain assumption Spatially adjacent Gaussians on the same object almost always share one semantic anchor, so the binding has low conditional entropy after Morton reordering.
    Stated in Introduction and exploited in Sec. 3.5 rate–distortion coding.
  • ad hoc to paper Decoupled transmittance-weighted least-squares surrogate is a sufficient substitute for the coupled rendering objective for per-Gaussian features.
    Sec. 3.3 replaces the full Gram-coupled objective with the closed-form average; accuracy is recovered by multi-view denoising rather than by proving equivalence.
invented entities (2)
  • Spatially grounded semantic anchors with joint spatial–semantic binding no independent evidence
    purpose: Replace per-Gaussian high-dim features by a small table plus assignment that is both accurate and compressible.
    Core construct of Sec. 3.3; evidence is only the paper’s own benchmarks.
  • Hierarchical parent-pointer binding coder (Morton + cross-level mode tables) no independent evidence
    purpose: Drive multi-level binding storage to a few bits per Gaussian without shipping a learned decoder.
    Sec. 3.5; compression gains shown only in Table 7 of this paper.

pith-pipeline@v1.1.0-grok45 · 20178 in / 3286 out tokens · 41678 ms · 2026-07-14T13:18:27.226358+00:00 · methodology

0 comments
read the original abstract

Open-vocabulary 3D scene understanding is commonly achieved by embedding 2D vision-language features such as CLIP into a 3D Gaussian Splatting scene, turning it into a text-queryable semantic field. However, attaching a high-dimensional feature to each of millions of Gaussians inflates a single scene to gigabytes, which makes storage and deployment the real bottleneck of these fields. Existing compact methods each learn and ship a per-scene codec, an autoencoder, a quantized codebook, or a distilled feature field, entangling field construction with field storage and never compressing the per-Gaussian assignment that holds the bulk of the cost. We argue that construction and storage should be decoupled, and that storage is a rate-distortion problem over the per-Gaussian binding to a small anchor table, a structure no prior open-vocabulary method compresses. We present CoSAG, which constructs the field without any per-scene training through a closed-form transmittance-weighted lift, spatially grounded semantic anchors, and multi-view denoising, and stores it with a spatially predictive entropy coder that ships no decoder. Because the anchors are spatially grounded, the binding is predictable and therefore highly compressible. The transmittance-weighted lift and multi-view denoising yield a clean, view-consistent assignment, so the entropy coder spends almost no rate on correcting noise and instead codes only the residual against its spatial prediction. CoSAG reaches sub-megabyte storage while matching or exceeding the state of the art across the 2D-rendered, 3D-selection, and dense-LSeg protocols, reducing field size by 37 to 76x relative to LangSplatV2 at higher accuracy.

Figures

Figures reproduced from arXiv: 2607.10237 by Jinlong Wang, Junhong Lin, Ruiting Dai, Wei Gao, Yuang Jia.

Figure 1
Figure 1. Figure 1: Compressing an open-vocabulary semantic field on the same pretrained 3D Gaussians. (a) Autoencoders suffer bottleneck loss and slow decoding. (b) Codebook indexing incurs extra training and quantization loss. (c) Additional-field methods train a separate semantic field, leaving two disjoint Gaussian fields whose appearance and semantics must be rendered separately. Our CoSAG (d) is training-free: it binds … view at source ↗
Figure 2
Figure 2. Figure 2: Overview of CoSAG. We lift per-pixel CLIP features over SAM regions in closed form onto a frozen 3DGS to obtain per-Gaussian features. Each Gaussian is then bound to a spatially grounded anchor by joint spatial and semantic proximity, and the anchors are denoised across views into three granularities. For storage, a shared basis compresses the anchor table while a spatially predictive coder handles the bin… view at source ↗
Figure 3
Figure 3. Figure 3: Qualitative 3D object selection on LERF-OVS. CoSAG selects the Gaussians of the queried object with coherent boundaries, while competing methods leak into surrounding geometry or miss parts of the object [PITH_FULL_IMAGE:figures/full_fig_p007_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: Storage–accuracy trade-off of the projection di￾mension. A uniform D (blue) saturates near 64 and collapses below it, while our per-level allocation [16, 32, 128] (orange) is Pareto-optimal and gains about +6 mIoU at equal storage. editing such as object removal, recoloring, and duplica￾tion. Additional localization, Replica, editing, and large-scale Waymo (Sun et al. 2020) results appear in the appendix. … view at source ↗
Figure 6
Figure 6. Figure 6: Qualitative 2D localization on LERF. CoSAG places a tighter, more focused relevancy peak on the queried object than the baselines, with fewer spurious activations [PITH_FULL_IMAGE:figures/full_fig_p011_6.png] view at source ↗
Figure 7
Figure 7. Figure 7: 3D scene editing enabled by CoSAG. A text query selects an object in 3D and supports removal, recoloring, and duplication, with the edited scene staying geometrically consistent [PITH_FULL_IMAGE:figures/full_fig_p012_7.png] view at source ↗
Figure 8
Figure 8. Figure 8 [PITH_FULL_IMAGE:figures/full_fig_p013_8.png] view at source ↗
Figure 9
Figure 9. Figure 9: Dense semantic segmentation on Replica room_0. Under the LSeg protocol, CoSAG reconstructs the dense semantic field faithfully while storing it in only 0.34 MB [PITH_FULL_IMAGE:figures/full_fig_p014_9.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

67 extracted references · 6 linked inside Pith

  1. [1]

    Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , pages=

    Langsplat: 3d language gaussian splatting , author=. Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , pages=

  2. [2]

    Advances in Neural Information Processing Systems , volume=

    Langsplatv2: High-dimensional 3d language gaussian splatting with 450+ fps , author=. Advances in Neural Information Processing Systems , volume=

  3. [3]

    Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , pages=

    Language embedded 3d gaussians for open-vocabulary scene understanding , author=. Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , pages=

  4. [4]

    Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , pages=

    Feature 3dgs: Supercharging 3d gaussian splatting to enable distilled feature fields , author=. Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , pages=

  5. [5]

    Advances in Neural Information Processing Systems , volume=

    Opengaussian: Towards point-level 3d gaussian-based open vocabulary understanding , author=. Advances in Neural Information Processing Systems , volume=

  6. [6]

    European Conference on Computer Vision , pages=

    N2f2: Hierarchical scene understanding with nested neural feature fields , author=. European Conference on Computer Vision , pages=. 2024 , organization=

    2024

  7. [7]

    European conference on computer vision , pages=

    Gaussian grouping: Segment and edit anything in 3d scenes , author=. European conference on computer vision , pages=. 2024 , organization=

    2024

  8. [8]

    Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , pages=

    LightSplat: Fast and Memory-Efficient Open-Vocabulary 3D Scene Understanding in Five Seconds , author=. Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , pages=

  9. [9]

    Proceedings of the IEEE/CVF conference on computer vision and pattern recognition , pages=

    Scaffold-gs: Structured 3d gaussians for view-adaptive rendering , author=. Proceedings of the IEEE/CVF conference on computer vision and pattern recognition , pages=

  10. [10]

    European Conference on Computer Vision , pages=

    Hac: Hash-grid assisted context for 3d gaussian splatting compression , author=. European Conference on Computer Vision , pages=. 2024 , organization=

    2024

  11. [11]

    IEEE Transactions on Pattern Analysis and Machine Intelligence , year=

    Hac++: Towards 100x compression of 3d gaussian splatting , author=. IEEE Transactions on Pattern Analysis and Machine Intelligence , year=

  12. [12]

    Advances in neural information processing systems , volume=

    Contextgs: Compact 3d gaussian splatting with anchor level context model , author=. Advances in neural information processing systems , volume=

  13. [13]

    arXiv preprint arXiv:2603.28431 , year=

    LG-HCC: Local Geometry-Aware Hierarchical Context Compression for 3D Gaussian Splatting , author=. arXiv preprint arXiv:2603.28431 , year=

    arXiv

  14. [14]

    ICASSP 2026-2026 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP) , pages=

    EntropyGS: an efficient entropy coding on 3D Gaussian splatting , author=. ICASSP 2026-2026 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP) , pages=. 2026 , organization=

    2026

  15. [15]

    arXiv preprint arXiv:2601.12814 , year=

    CSGaussian: Progressive Rate-Distortion Compression and Segmentation for 3D Gaussian Splatting , author=. arXiv preprint arXiv:2601.12814 , year=

    arXiv

  16. [16]

    , author=

    3d gaussian splatting for real-time radiance field rendering. , author=. ACM Trans. Graph. , volume=

  17. [17]

    International conference on machine learning , pages=

    Learning transferable visual models from natural language supervision , author=. International conference on machine learning , pages=. 2021 , organization=

    2021

  18. [18]

    Proceedings of the IEEE/CVF international conference on computer vision , pages=

    Segment anything , author=. Proceedings of the IEEE/CVF international conference on computer vision , pages=

  19. [19]

    Proceedings of the IEEE/CVF international conference on computer vision , pages=

    Lerf: Language embedded radiance fields , author=. Proceedings of the IEEE/CVF international conference on computer vision , pages=

  20. [22]

    Proceedings of the IEEE/CVF conference on computer vision and pattern recognition , pages=

    Openscene: 3d scene understanding with open vocabularies , author=. Proceedings of the IEEE/CVF conference on computer vision and pattern recognition , pages=

  21. [24]

    Presence: teleoperators & virtual environments , volume=

    A survey of augmented reality , author=. Presence: teleoperators & virtual environments , volume=. 1997 , publisher=

    1997

  22. [25]

    splat: Directly referring 3d gaussian splatting via direct language embedding registration , author=

    Dr. splat: Directly referring 3d gaussian splatting via direct language embedding registration , author=. Proceedings of the Computer Vision and Pattern Recognition Conference , pages=

  23. [26]

    Advances in Neural Information Processing Systems , volume=

    Ea3d: Online open-world 3d object extraction from streaming videos , author=. Advances in Neural Information Processing Systems , volume=

  24. [28]

    Advances in Neural Information Processing Systems , volume=

    Weakly supervised 3d open-vocabulary segmentation , author=. Advances in Neural Information Processing Systems , volume=

  25. [29]

    Proceedings of the Computer Vision and Pattern Recognition Conference , pages=

    Efficient decoupled feature 3d gaussian splatting via hierarchical compression , author=. Proceedings of the Computer Vision and Pattern Recognition Conference , pages=

  26. [31]

    Proceedings of the AAAI Conference on Artificial Intelligence , number=

    Gags: Granularity-aware feature distillation for language gaussian splatting , author=. Proceedings of the AAAI Conference on Artificial Intelligence , number=

  27. [33]

    International Journal of Computer Vision , volume=

    Fmgs: Foundation model embedded 3d gaussian splatting for holistic 3d scene understanding , author=. International Journal of Computer Vision , volume=. 2025 , publisher=

    2025

  28. [34]

    IEEE Transactions on Pattern Analysis and Machine Intelligence , year=

    Langsurf: Language-embedded surface gaussians for 3d scene understanding , author=. IEEE Transactions on Pattern Analysis and Machine Intelligence , year=

  29. [35]

    Proceedings of the IEEE/CVF International Conference on Computer Vision , pages=

    CLIP-GS: Unifying vision-language representation with 3D Gaussian splatting , author=. Proceedings of the IEEE/CVF International Conference on Computer Vision , pages=

  30. [36]

    Proceedings of the Computer Vision and Pattern Recognition Conference , pages=

    Instancegaussian: Appearance-semantic joint gaussian representation for 3d instance-level perception , author=. Proceedings of the Computer Vision and Pattern Recognition Conference , pages=

  31. [37]

    Proceedings of the IEEE/CVF International Conference on Computer Vision , pages=

    In-place scene labelling and understanding with implicit scene representation , author=. Proceedings of the IEEE/CVF International Conference on Computer Vision , pages=

  32. [38]

    Proceedings of the IEEE/CVF International Conference on Computer Vision , pages=

    Ludvig: Learning-free uplifting of 2d visual features to gaussian splatting scenes , author=. Proceedings of the IEEE/CVF International Conference on Computer Vision , pages=

  33. [39]

    Proceedings of the IEEE/CVF conference on computer vision and pattern recognition , pages=

    Scalability in perception for autonomous driving: Waymo open dataset , author=. Proceedings of the IEEE/CVF conference on computer vision and pattern recognition , pages=

  34. [40]

    Azuma, R. T. 1997. A survey of augmented reality. Presence: teleoperators & virtual environments, 6(4): 355--385

    1997

  35. [41]

    Bang, J.; Kim, J.; Kim, M.; Jeong, S.; and Joo, K. 2026. LightSplat: Fast and Memory-Efficient Open-Vocabulary 3D Scene Understanding in Five Seconds. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 19812--19821

    2026

  36. [42]

    A.; Guan, Y.; Ryhner, S.; Lon c ari \'c , S.; and Vijaykumar, N

    Budimir, L. A.; Guan, Y.; Ryhner, S.; Lon c ari \'c , S.; and Vijaykumar, N. 2026. Sparse Code Uplifting for Efficient 3D Language Gaussian Splatting. arXiv preprint arXiv:2605.13600

    Pith/arXiv arXiv 2026

  37. [43]

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

    2024

  38. [44]

    Dai, Z.; Liu, T.; and Zhang, Y. 2025. Efficient decoupled feature 3d gaussian splatting via hierarchical compression. In Proceedings of the Computer Vision and Pattern Recognition Conference, 11156--11166

    2025

  39. [45]

    Deng, T.; Chen, X.; Chen, Y.; Chen, Q.; Xu, Y.; Yang, L.; Xu, L.; Zhang, Y.; Zhang, B.; Huang, W.; et al. 2025. GaussianDWM: 3D Gaussian Driving World Model for Unified Scene Understanding and Multi-Modal Generation. arXiv preprint arXiv:2512.23180

    Pith/arXiv arXiv 2025

  40. [46]

    M.; Kuwajerwala, A.; Gu, Q.; Omama, M.; Chen, T.; Maalouf, A.; Li, S.; Iyer, G.; Saryazdi, S.; Keetha, N.; et al

    Jatavallabhula, K. M.; Kuwajerwala, A.; Gu, Q.; Omama, M.; Chen, T.; Maalouf, A.; Li, S.; Iyer, G.; Saryazdi, S.; Keetha, N.; et al. 2023. Conceptfusion: Open-set multimodal 3d mapping. arXiv preprint arXiv:2302.07241

    Pith/arXiv arXiv 2023

  41. [47]

    Jiao, S.; Dong, H.; Yin, Y.; Jie, Z.; Qian, Y.; Zhao, Y.; Shi, H.; and Wei, Y. 2025. CLIP-GS: Unifying vision-language representation with 3D Gaussian splatting. In Proceedings of the IEEE/CVF International Conference on Computer Vision, 4670--4680

    2025

  42. [48]

    F.; Choe, J.; and Oh, T.-H

    Jun-Seong, K.; Kim, G.; Yu-Ji, K.; Wang, Y.-C. F.; Choe, J.; and Oh, T.-H. 2025. Dr. splat: Directly referring 3d gaussian splatting via direct language embedding registration. In Proceedings of the Computer Vision and Pattern Recognition Conference, 14137--14146

    2025

  43. [49]

    Kerbl, B.; Kopanas, G.; Leimk \"u hler, T.; Drettakis, G.; et al. 2023. 3d gaussian splatting for real-time radiance field rendering. ACM Trans. Graph., 42(4): 139--1

    2023

  44. [50]

    M.; Goldberg, K.; Kanazawa, A.; and Tancik, M

    Kerr, J.; Kim, C. M.; Goldberg, K.; Kanazawa, A.; and Tancik, M. 2023. Lerf: Language embedded radiance fields. In Proceedings of the IEEE/CVF international conference on computer vision, 19729--19739

    2023

  45. [51]

    C.; Lo, W.-Y.; et al

    Kirillov, A.; Mintun, E.; Ravi, N.; Mao, H.; Rolland, C.; Gustafson, L.; Xiao, T.; Whitehead, S.; Berg, A. C.; Lo, W.-Y.; et al. 2023. Segment anything. In Proceedings of the IEEE/CVF international conference on computer vision, 4015--4026

    2023

  46. [52]

    Lee, H.; Min, J.; and Park, J. 2025. CF3: Compact and Fast 3D Feature Fields. arXiv preprint arXiv:2508.05254

    Pith/arXiv arXiv 2025

  47. [53]

    Q.; Belongie, S.; Koltun, V.; and Ranftl, R

    Li, B.; Weinberger, K. Q.; Belongie, S.; Koltun, V.; and Ranftl, R. 2022. Language-driven semantic segmentation. arXiv preprint arXiv:2201.03546

    Pith/arXiv arXiv 2022

  48. [54]

    Li, H.; Qin, M.; Zou, Z.; He, D.; Ji, X.; Li, B.; Dai, B.; Zhang, D.; and Han, J. 2026 a . Langsurf: Language-embedded surface gaussians for 3d scene understanding. IEEE Transactions on Pattern Analysis and Machine Intelligence

    2026

  49. [55]

    Li, H.; Wu, Y.; Meng, J.; Gao, Q.; Zhang, Z.; Wang, R.; and Zhang, J. 2025. Instancegaussian: Appearance-semantic joint gaussian representation for 3d instance-level perception. In Proceedings of the Computer Vision and Pattern Recognition Conference, 14078--14088

    2025

  50. [56]

    Li, W.; Zhao, Y.; Qin, M.; Liu, Y.; Cai, Y.; Gan, C.; and Pfister, H. 2026 b . Langsplatv2: High-dimensional 3d language gaussian splatting with 450+ fps. Advances in Neural Information Processing Systems, 38: 174306--174330

    2026

  51. [57]

    Liu, K.; Zhan, F.; Zhang, J.; Xu, M.; Yu, Y.; El Saddik, A.; Theobalt, C.; Xing, E.; and Lu, S. 2023. Weakly supervised 3d open-vocabulary segmentation. Advances in Neural Information Processing Systems, 36: 53433--53456

    2023

  52. [58]

    Lu, T.; Yu, M.; Xu, L.; Xiangli, Y.; Wang, L.; Lin, D.; and Dai, B. 2024. Scaffold-gs: Structured 3d gaussians for view-adaptive rendering. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, 20654--20664

    2024

  53. [59]

    Marrie, J.; M \'e n \'e gaux, R.; Arbel, M.; Larlus, D.; and Mairal, J. 2025. Ludvig: Learning-free uplifting of 2d visual features to gaussian splatting scenes. In Proceedings of the IEEE/CVF International Conference on Computer Vision, 7440--7450

    2025

  54. [60]

    Peng, S.; Genova, K.; Jiang, C.; Tagliasacchi, A.; Pollefeys, M.; Funkhouser, T.; et al. 2023. Openscene: 3d scene understanding with open vocabularies. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, 815--824

    2023

  55. [61]

    Peng, Y.; Wang, H.; Liu, Y.; Wen, C.; Dong, Z.; and Yang, B. 2026. Gags: Granularity-aware feature distillation for language gaussian splatting. In Proceedings of the AAAI Conference on Artificial Intelligence, 10, 8376--8384

    2026

  56. [62]

    Qin, M.; Li, W.; Zhou, J.; Wang, H.; and Pfister, H. 2024. Langsplat: 3d language gaussian splatting. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 20051--20060

    2024

  57. [63]

    W.; Hallacy, C.; Ramesh, A.; Goh, G.; Agarwal, S.; Sastry, G.; Askell, A.; Mishkin, P.; Clark, J.; et al

    Radford, A.; Kim, J. W.; Hallacy, C.; Ramesh, A.; Goh, G.; Agarwal, S.; Sastry, G.; Askell, A.; Mishkin, P.; Clark, J.; et al. 2021. Learning transferable visual models from natural language supervision. In International conference on machine learning, 8748--8763. PmLR

    2021

  58. [64]

    Shi, J.-C.; Wang, M.; Duan, H.-B.; and Guan, S.-H. 2024. Language embedded 3d gaussians for open-vocabulary scene understanding. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 5333--5343

    2024

  59. [65]

    J.; Mur-Artal, R.; Ren, C.; Verma, S.; et al

    Straub, J.; Whelan, T.; Ma, L.; Chen, Y.; Wijmans, E.; Green, S.; Engel, J. J.; Mur-Artal, R.; Ren, C.; Verma, S.; et al. 2019. The replica dataset: A digital replica of indoor spaces. arXiv preprint arXiv:1906.05797

    Pith/arXiv arXiv 2019

  60. [66]

    Sun, P.; Kretzschmar, H.; Dotiwalla, X.; Chouard, A.; Patnaik, V.; Tsui, P.; Guo, J.; Zhou, Y.; Chai, Y.; Caine, B.; et al. 2020. Scalability in perception for autonomous driving: Waymo open dataset. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, 2446--2454

    2020

  61. [67]

    C.; and Wen, B

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

    2024

  62. [68]

    Wu, Y.; Meng, J.; Li, H.; Wu, C.; Shi, Y.; Cheng, X.; Zhao, C.; Feng, H.; Ding, E.; Wang, J.; et al. 2024. Opengaussian: Towards point-level 3d gaussian-based open vocabulary understanding. Advances in Neural Information Processing Systems, 37: 19114--19138

    2024

  63. [69]

    Ye, M.; Danelljan, M.; Yu, F.; and Ke, L. 2024. Gaussian grouping: Segment and edit anything in 3d scenes. In European conference on computer vision, 162--179. Springer

    2024

  64. [70]

    Zhi, S.; Laidlow, T.; Leutenegger, S.; and Davison, A. J. 2021. In-place scene labelling and understanding with implicit scene representation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, 15838--15847

    2021

  65. [71]

    Zhou, S.; Chang, H.; Jiang, S.; Fan, Z.; Zhu, Z.; Xu, D.; Chari, P.; You, S.; Wang, Z.; and Kadambi, A. 2024. Feature 3dgs: Supercharging 3d gaussian splatting to enable distilled feature fields. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 21676--21685

    2024

  66. [72]

    Zhou, X.; Wang, J.; Jia, Y.; Wang, Y.; Sun, D.; and Yang, M.-H. 2026. Ea3d: Online open-world 3d object extraction from streaming videos. Advances in Neural Information Processing Systems, 38: 54805--54825

    2026

  67. [73]

    Zuo, X.; Samangouei, P.; Zhou, Y.; Di, Y.; and Li, M. 2025. Fmgs: Foundation model embedded 3d gaussian splatting for holistic 3d scene understanding. International Journal of Computer Vision, 133(2): 611--627

    2025