Pith. sign in

REVIEW 2 major objections 45 references

Multi-cue mask refinement turns inconsistent 2D segments into coherent 3D instance labels for Gaussian Splatting scenes.

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-4.3

2026-07-03 16:48 UTC pith:MEUFWGNU

load-bearing objection The paper outlines a multi-cue pipeline to reduce mask fragmentation when lifting 2D segments into 3D Gaussian Splatting, but the abstract supplies no metrics to show whether the gains are real. the 2 major comments →

arxiv 2607.01708 v1 pith:MEUFWGNU submitted 2026-07-02 cs.CV

Consistent Scene Understanding in 3D Gaussian Splatting via Multi-Cue Mask Refinement

Hyunjoon Park , Donghyeon Cho This is my paper
classification cs.CV
keywords 3D Gaussian Splattinginstance segmentationmask refinementcross-view consistencyscene understanding3D reconstructionmulti-cue merging
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.

The paper seeks to solve the fragmentation and view-to-view inconsistency that arise when 2D foundation segmentation models are applied to 3D Gaussian Splatting. It introduces a three-stage pipeline that first pulls semantic, geometric, and structural cues from each image, then merges broken masks with a composite score built from those cues, and finally assigns the same identity to matching regions across all viewpoints. If the pipeline works, viewpoint-specific fragments become stable 3D object primitives that support reliable instance segmentation and object-level editing while the underlying photometric reconstruction stays accurate. Readers would care because consistent object identities are a basic requirement for any downstream task that treats the 3D scene as a collection of editable parts rather than a single image.

Core claim

The central claim is that extracting synergistic semantic, geometric, and structural priors, consolidating fragmented masks through a composite merge score derived from semantic, depth, and edge cues, and performing cross-view mask matching to assign globally consistent identities together transform viewpoint-specific segments into coherent 3D primitives, thereby enabling stable 3D instance segmentation and effective downstream editing tasks in 3D Gaussian Splatting.

What carries the argument

The multi-cue-guided mask merging process that consolidates fragmented masks using a composite merge score from semantic, depth, and edge cues, together with the cross-view mask matching step that establishes globally consistent identity assignments.

Load-bearing premise

That cues extracted from separate 2D foundation models can be fused into a single composite score and matched across views to produce consistent identities without creating new fragments or identity switches.

What would settle it

A quantitative evaluation on a held-out set of scenes in which the method records lower cross-view consistency scores or higher rates of identity switches than the raw 2D model outputs would show the refinement steps do not deliver the claimed improvement.

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

If this is right

  • Stable 3D instance segmentation becomes available across multiple viewpoints.
  • Downstream object-level editing tasks can operate on coherent 3D primitives rather than fragmented 2D masks.
  • Cross-view consistency and segmentation stability improve relative to baselines that use 2D masks directly.
  • High-fidelity photometric reconstruction quality is preserved during the added segmentation steps.

Where Pith is reading between the lines

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

  • The same cue-merging logic could be tested on other neural rendering backbones to check whether the consistency gains transfer beyond Gaussian Splatting.
  • If the composite score generalizes, the pipeline might reduce the amount of manual mask cleanup needed in production 3D reconstruction workflows.
  • Consistent 3D identities could support new tasks such as per-object material editing or physics simulation inside the reconstructed scene.
  • The approach implicitly assumes static scenes; applying it to video with moving objects would require an additional temporal matching stage.

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 / 0 minor

Summary. The manuscript proposes a three-stage framework to produce consistent 2D instance masks that guide optimization of 3D Gaussian Splatting feature fields. Stage (1) performs Multi-Cue Extraction to obtain synergistic semantic, geometric, and structural priors from 2D foundation models on input images. Stage (2) applies Multi-Cue-Guided Mask Merging that consolidates fragmented masks via a composite merge score derived from semantic, depth, and edge cues. Stage (3) performs Cross-View Mask Matching to establish globally consistent identity assignments across viewpoints. The central claim is that this transforms viewpoint-specific segments into coherent 3D primitives, yielding stable 3D instance segmentation and effective downstream editing while preserving high-fidelity photometric reconstruction, with experiments showing significant gains in cross-view consistency and segmentation stability over baselines.

Significance. If the central claim holds with supporting evidence, the work would address a practical limitation in lifting inconsistent 2D segmentation priors to coherent 3D representations, which is relevant for object-level scene understanding and editing tasks in neural rendering pipelines. The multi-cue approach to merging and matching is a plausible direction for reducing fragmentation without new identity errors. However, the absence of any quantitative metrics, ablation studies, error analysis, or implementation details prevents assessment of whether the improvements are substantial, robust, or generalizable.

major comments (2)
  1. [Abstract] Abstract: the assertion that the method 'significantly improves cross-view consistency and segmentation stability over existing baselines' is load-bearing for the central claim yet is unsupported by any quantitative metrics, ablation studies, error analysis, or implementation details, making verification of the framework's effectiveness impossible from the supplied information.
  2. [Abstract] Abstract, stage (2): the composite merge score is presented as the mechanism for consolidating fragmented masks, but no formulation, weighting scheme, or matching procedure across views is supplied; this is central to the weakest assumption that synergistic priors can be merged without introducing new fragmentation or identity errors.

Simulated Author's Rebuttal

2 responses · 0 unresolved

We thank the referee for the constructive feedback on the abstract. We agree that the central claims require clearer quantitative grounding and explicit method details to be verifiable. We will revise the abstract accordingly while preserving its concise nature.

read point-by-point responses

  1. Referee: [Abstract] Abstract: the assertion that the method 'significantly improves cross-view consistency and segmentation stability over existing baselines' is load-bearing for the central claim yet is unsupported by any quantitative metrics, ablation studies, error analysis, or implementation details, making verification of the framework's effectiveness impossible from the supplied information.

    Authors: We agree that the abstract claim would be stronger with direct quantitative support. In the revised manuscript we will incorporate specific metrics (e.g., cross-view consistency IoU gains and segmentation stability scores) into the abstract, together with a brief reference to the ablation and error analysis sections, so the improvement statement can be assessed from the abstract alone. revision: yes

  2. Referee: [Abstract] Abstract, stage (2): the composite merge score is presented as the mechanism for consolidating fragmented masks, but no formulation, weighting scheme, or matching procedure across views is supplied; this is central to the weakest assumption that synergistic priors can be merged without introducing new fragmentation or identity errors.

    Authors: We acknowledge that the abstract omits the mathematical formulation. We will revise the abstract to include a concise description of the composite merge score (weighted sum of semantic, depth, and edge terms) and note that the full weighting scheme and cross-view matching procedure appear in Section 3. This will make the merging mechanism explicit without exceeding abstract length limits. revision: partial

Circularity Check

0 steps flagged

No significant circularity; derivation is self-contained

full rationale

The paper presents a three-stage pipeline (multi-cue extraction, mask merging via composite score, cross-view matching) for lifting 2D segments to consistent 3D primitives in Gaussian Splatting. No equations, fitted parameters, or first-principles derivations appear in the abstract or described framework. The central claim rests on empirical merging and matching heuristics whose correctness is evaluated externally via consistency metrics and editing tasks, not by construction from the inputs themselves. No self-citation chains, ansatzes, or renamings reduce the output to the input. This is the normal case of an applied method paper whose load-bearing steps are algorithmic rather than tautological.

Axiom & Free-Parameter Ledger

0 free parameters · 0 axioms · 0 invented entities

Abstract provides no explicit free parameters, axioms, or invented entities; the method is described as relying on existing 2D foundation segmentation models and standard 3DGS optimization.

pith-pipeline@v0.9.1-grok · 5730 in / 1068 out tokens · 35692 ms · 2026-07-03T16:48:38.921419+00:00 · methodology

0 comments
read the original abstract

Reliable instance-level scene understanding is a fundamental prerequisite for object-level interactions and high-fidelity 3D representations. While current methods often leverage 2D foundation segmentation models to obtain these priors, their 2D-centric design typically yields fragmented masks and inconsistent predictions across different views. To address these issues, we propose a novel framework that produces consistent 2D instance masks to guide the optimization of 3D Gaussian Splatting (3DGS) feature fields. Our framework consists of three main stages. (1) Multi-Cue Extraction that generates synergistic semantic, geometric, and structural priors from input images. (2) Multi-Cue-Guided Mask Merging process that consolidates fragmented masks using a composite merge score derived from semantic, depth, and edge cues. (3) Cross-View Mask Matching that establishes globally consistent identity assignments across all viewpoints. By transforming viewpoint-specific segments into coherent 3D primitives, our approach enables stable 3D instance segmentation and effective downstream editing tasks. Experiments demonstrate that our method significantly improves cross-view consistency and segmentation stability over existing baselines while maintaining high-fidelity photometric reconstruction.

Figures

Figures reproduced from arXiv: 2607.01708 by Donghyeon Cho, Hyunjoon Park.

Figure 1
Figure 1. Figure 1: Cross-view inconsistency from SAM over-segmentation. Auto￾matic masks from SAM [12] suffer from inconsistent identities across distant (N . . . M) and even contiguous frames (K ± 1). Its 2D-centric nature induces over-segmentation and erroneous merging of distinct objects, impeding coherent 3D scene understanding. these models into 3DGS remains challenging because 2D models lack 3D spa￾tial awareness and g… view at source ↗
Figure 2
Figure 2. Figure 2: Pipeline. Given multiview images, we extract initial SAM masks and multi-cue priors. These masks are refined via MCM using a composite merge score. Following cross-view mask matching, refined 2D features are back￾projected and lifted into 3D Gaussians. Consequently, the model outputs a view￾consistent object-level feature field for high-fidelity 3D scene understanding. mean depth ¯di and depth gradient mag… view at source ↗
Figure 3
Figure 3. Figure 3: Qualitative Result. This figure demonstrates how faithfully our results represent the scene and maintain view consistency across different viewpoints. 5.3 Quantitative Results [PITH_FULL_IMAGE:figures/full_fig_p011_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: Qualitative Comparison of View Consistency. Compared to the baseline [40], which exhibits over-segmentation and inconsistent object identi￾ties across different views, our approach successfully produces coherent, view￾consistent object masks across changing viewpoints [PITH_FULL_IMAGE:figures/full_fig_p012_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: Qualitative Ablation of Cross-View Mask Matching. Without cross-view matching (middle row), masks are restricted to per-view assignments, leading to inconsistent identities. Our full pipeline (bottom row) ensures global identity coherence, which is essential for stable 3D representation. metrics through precise boundary delineation. Our full model, which includes DepthAnythingV2 geometric priors, achieves … view at source ↗
Figure 6
Figure 6. Figure 6: Visualization of multi-cue features. We extract complementary pri￾ors to address the limitations of the 2D-centric proposals of SAM. While the Initial SAM Mask exhibits severe oversegmentation, our framework leverages Depth Maps (geometric context), Edge Maps (structural discontinuities), and Semantics (high-level object coherence) to consolidate fragmented regions. This multi-cue integration ensures that … view at source ↗
Figure 7
Figure 7. Figure 7: Additional Qualitative Result. Our method successfully generates globally coherent 3D instance masks across various indoor and real-world scenes. 8.4 Synthesis of Multi-Cues The integration of these three models addresses the fundamental limitations of single-modality reasoning. While depth resolves appearance aliasing, seman￾tics capture high-level coherence, and edge detection identifies boundaries that … view at source ↗
Figure 8
Figure 8. Figure 8: AI generated Images. We utilized these images in our main pipeline to illustrate effectively. – Prompt 1 (Object Isolation): "An illustration showing a target object (an elephant) isolated from a com￾plex scene using a segmentation mask, with multiple camera frustums from different angles pointing at the object to represent a multiview capture setup." – Prompt 2 (Feature Lifting): "A conceptual diagram sho… 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

45 extracted references · 45 canonical work pages · 1 internal anchor

  1. [1]

    In: ICCV

    Barron, J.T., Mildenhall, B., Tancik, M., Hedman, P., Martin-Brualla, R., Srini- vasan, P.P.: Mip-nerf: A multiscale representation for anti-aliasing neural radiance fields. In: ICCV. pp. 5855–5864 (2021) 15 22 Hyunjoon Park and Donghyeon Cho

    work page 2021

  2. [2]

    In: ECCV

    Bhalgat, Y., Laina, I., Henriques, J.F., Zisserman, A., Vedaldi, A.: N2f2: Hierarchi- cal scene understanding with nested neural feature fields. In: ECCV. pp. 197–214. Springer (2024) 3

    work page 2024

  3. [3]

    TPAMI (6), 679–698 (1986) 16

    Canny, J.: A computational approach to edge detection. TPAMI (6), 679–698 (1986) 16

    work page 1986

  4. [4]

    In: AAAI

    Cen, J., Fang, J., Yang, C., Xie, L., Zhang, X., Shen, W., Tian, Q.: Segment any 3d gaussians. In: AAAI. vol. 39, pp. 1971–1979 (2025) 3, 10, 11

    work page 1971

  5. [5]

    In: CVPR

    Chou, Z.T., Huang, S.Y., Liu, I., Wang, Y.C.F., et al.: Gsnerf: Generalizable se- mantic neural radiance fields with enhanced 3d scene understanding. In: CVPR. pp. 20806–20815 (2024) 3

    work page 2024

  6. [6]

    arXiv (2024) 3

    Guo, J., Ma, X., Fan, Y., Liu, H., Li, Q.: Semantic gaussians: Open-vocabulary scene understanding with 3d gaussian splatting. arXiv (2024) 3

    work page 2024

  7. [7]

    Hore, A., Ziou, D.: Image quality metrics: Psnr vs. ssim. In: ICPR. pp. 2366–2369. IEEE (2010) 9

    work page 2010

  8. [8]

    Hu,X.,Wang,Y.,Fan,L.,Luo,C.,Fan,J.,Lei,Z.,Li,Q.,Peng,J.,Zhang,Z.:Sagd: Boundary-enhanced segment anything in 3d gaussian via gaussian decomposition (2025), https://arxiv.org/abs/2401.17857 3

    work page arXiv 2025

  9. [9]

    ACM Trans

    Kerbl, B., Kopanas, G., Leimkühler, T., Drettakis, G.: 3d gaussian splatting for real-time radiance field rendering. ACM Trans. Graph.42(4), 139–1 (2023) 1, 3, 9, 10, 11

    work page 2023

  10. [10]

    In: ICCV

    Kerr, J., Kim, C.M., Goldberg, K., Kanazawa, A., Tancik, M.: Lerf: Language embedded radiance fields. In: ICCV. pp. 19729–19739 (2023) 3, 9, 15

    work page 2023

  11. [11]

    In: CVPR

    Kim, C.M., Wu, M., Kerr, J., Goldberg, K., Tancik, M., Kanazawa, A.: Garfield: Group anything with radiance fields. In: CVPR. pp. 21530–21539 (2024) 3, 10, 11

    work page 2024

  12. [12]

    In: ICCV

    Kirillov, A., Mintun, E., Ravi, N., Mao, H., Rolland, C., Gustafson, L., Xiao, T., Whitehead, S., Berg, A.C., Lo, W.Y., et al.: Segment anything. In: ICCV. pp. 4015–4026 (2023) 1, 2, 3, 4, 20

    work page 2023

  13. [14]

    NeurIPS35, 23311–23330 (2022) 3

    Kobayashi, S., Matsumoto, E., Sitzmann, V.: Decomposing nerf for editing via feature field distillation. NeurIPS35, 23311–23330 (2022) 3

    work page 2022

  14. [15]

    Naval research logistics quarterly2(1-2), 83–97 (1955) 19

    Kuhn, H.W.: The hungarian method for the assignment problem. Naval research logistics quarterly2(1-2), 83–97 (1955) 19

    work page 1955

  15. [16]

    In: CVPR

    Landrieu, L., Simonovsky, M.: Large-scale point cloud semantic segmentation with superpoint graphs. In: CVPR. pp. 4558–4567 (2018) 3

    work page 2018

  16. [17]

    ICCV (2025) 3, 10, 11

    Lee, H., Min, J., Park, J.: Cf3: Compact and fast 3d feature fields. ICCV (2025) 3, 10, 11

    work page 2025

  17. [18]

    In: ICPR

    Lee, Y., Ryu, J., Yoon, D., Cho, D.: Consistent object removal from masked neural radiance fields by estimating never-seen regions in all-views. In: ICPR. pp. 416–431. Springer (2024) 9

    work page 2024

  18. [19]

    Li, B., Weinberger, K.Q., Belongie, S., Koltun, V., Ranftl, R.: Language-driven semantic segmentation (2022), https://arxiv.org/abs/2201.03546 15

    work page internal anchor Pith review Pith/arXiv arXiv 2022

  19. [20]

    In: CVPR

    Li, H., Wu, Y., Meng, J., Gao, Q., Zhang, Z., Wang, R., Zhang, J.: Instance- gaussian: Appearance-semantic joint gaussian representation for 3d instance-level perception. In: CVPR. pp. 14078–14088 (2025) 3, 10, 11

    work page 2025

  20. [21]

    In: ICCV

    Liu, M., Uy, M.A., Xiang, D., Su, H., Fidler, S., Sharp, N., Gao, J.: Partfield: Learning 3d feature fields for part segmentation and beyond. In: ICCV. pp. 9704– 9715 (2025) 3

    work page 2025

  21. [22]

    In: ECCV

    Liu, S., Zeng, Z., Ren, T., Li, F., Zhang, H., Yang, J., Jiang, Q., Li, C., Yang, J., Su, H., et al.: Grounding dino: Marrying dino with grounded pre-training for open-set object detection. In: ECCV. pp. 38–55. Springer (2024) 1 Multi-Cue Mask Refinement in 3DGS 23

    work page 2024

  22. [23]

    In: NeurIPS 3

    Lu, Y., Zhou, Y., Qiao, Y., Song, C., Liang, T., Ma, J., Wang, H., Yin, Y.: Segment then splat: Unified 3d open-vocabulary segmentation via gaussian splatting. In: NeurIPS 3

    work page

  23. [24]

    In: ICCV

    Ma, Z., Yue, Y., Gkioxari, G.: Find any part in 3d. In: ICCV. pp. 7818–7827 (2025) 3

    work page 2025

  24. [25]

    Proceedings of the Royal Society of London

    Marr, D., Hildreth, E.: Theory of edge detection. Proceedings of the Royal Society of London. Series B. Biological Sciences207(1167), 187–217 (1980) 4, 5, 12, 15

    work page 1980

  25. [26]

    Commu- nications of the ACM65(1), 99–106 (2021) 1

    Mildenhall, B., Srinivasan, P.P., Tancik, M., Barron, J.T., Ramamoorthi, R., Ng, R.: Nerf: Representing scenes as neural radiance fields for view synthesis. Commu- nications of the ACM65(1), 99–106 (2021) 1

    work page 2021

  26. [27]

    arXiv (2023) 2, 4, 12, 14, 15

    Oquab, M., Darcet, T., Moutakanni, T., Vo, H., Szafraniec, M., Khalidov, V., Fernandez, P., Haziza, D., Massa, F., El-Nouby, A., et al.: Dinov2: Learning robust visual features without supervision. arXiv (2023) 2, 4, 12, 14, 15

    work page 2023

  27. [28]

    Scholarpedia4(2), 1883 (2009) 9

    Peterson, L.E.: K-nearest neighbor. Scholarpedia4(2), 1883 (2009) 9

    work page 2009

  28. [29]

    In: CVPR

    Qin, M., Li, W., Zhou, J., Wang, H., Pfister, H.: Langsplat: 3d language gaussian splatting. In: CVPR. pp. 20051–20060 (2024) 3, 20

    work page 2024

  29. [30]

    In: ICML

    Radford, A., Kim, J.W., Hallacy, C., Ramesh, A., Goh, G., Agarwal, S., Sastry, G., Askell, A., Mishkin, P., Clark, J., et al.: Learning transferable visual models from natural language supervision. In: ICML. pp. 8748–8763. PmLR (2021) 1, 15, 20

    work page 2021

  30. [31]

    In: ICCV

    Ranftl, R., Bochkovskiy, A., Koltun, V.: Vision transformers for dense prediction. In: ICCV. pp. 12179–12188 (2021) 15

    work page 2021

  31. [32]

    TPAMI44(3), 1623–1637 (2020) 15

    Ranftl, R., Lasinger, K., Hafner, D., Schindler, K., Koltun, V.: Towards robust monocular depth estimation: Mixing datasets for zero-shot cross-dataset transfer. TPAMI44(3), 1623–1637 (2020) 15

    work page 2020

  32. [33]

    In: CVPR

    Shi, J.C., Wang, M., Duan, H.B., Guan, S.H.: Language embedded 3d gaussians for open-vocabulary scene understanding. In: CVPR. pp. 5333–5343 (2024) 3

    work page 2024

  33. [34]

    Presentation at Stanford Artificial Intelligence Project pp

    Sobel, I., Feldman, G.: An isotropic 3x3 gradient operator for image processing. Presentation at Stanford Artificial Intelligence Project pp. 271–272 (1968) 16

    work page 1968

  34. [35]

    arXiv (2019) 3, 9, 18

    Straub, J., Whelan, T., Ma, L., Chen, Y., Wijmans, E., Green, S., Engel, J.J., Mur-Artal, R., Ren, C., Verma, S., et al.: The replica dataset: A digital replica of indoor spaces. arXiv (2019) 3, 9, 18

    work page 2019

  35. [36]

    Tschernezki, V., Laina, I., Larlus, D., Vedaldi, A.: Neural feature fusion fields: 3d distillation of self-supervised 2d image representations. In: 3DV. pp. 443–453. IEEE (2022) 3

    work page 2022

  36. [37]

    In: ICASSP

    Wang, N., Yan, X., Song, X., Wang, Z.: Semantic-guided gaussian splatting with deferred rendering. In: ICASSP. pp. 1–5. IEEE (2025) 3

    work page 2025

  37. [38]

    NeurIPS37, 19114–19138 (2024) 3

    Wu, Y., Meng, J., Li, H., Wu, C., Shi, Y., Cheng, X., Zhao, C., Feng, H., Ding, E., Wang, J., et al.: Opengaussian: Towards point-level 3d gaussian-based open vocabulary understanding. NeurIPS37, 19114–19138 (2024) 3

    work page 2024

  38. [39]

    NeurIPS37, 21875–21911 (2024) 2, 4, 12, 15

    Yang, L., Kang, B., Huang, Z., Zhao, Z., Xu, X., Feng, J., Zhao, H.: Depth anything v2. NeurIPS37, 21875–21911 (2024) 2, 4, 12, 15

    work page 2024

  39. [40]

    In: ECCV

    Ye, M., Danelljan, M., Yu, F., Ke, L.: Gaussian grouping: Segment and edit any- thing in 3d scenes. In: ECCV. pp. 162–179. Springer (2024) 3, 9, 10, 11, 12, 20

    work page 2024

  40. [41]

    In: CVPR

    Yin, Y., Liu, Y., Xiao, Y., Cohen-Or, D., Huang, J., Chen, B.: Sai3d: Segment any instance in 3d scenes. In: CVPR. pp. 3292–3302 (2024) 3

    work page 2024

  41. [42]

    In: ECCV

    Yue, Y., Das, A., Engelmann, F., Tang, S., Lenssen, J.E.: Improving 2d feature representations by 3d-aware fine-tuning. In: ECCV. pp. 57–74. Springer (2024) 3

    work page 2024

  42. [43]

    In: CVPR

    Zhang, R., Isola, P., Efros, A.A., Shechtman, E., Wang, O.: The unreasonable effectiveness of deep features as a perceptual metric. In: CVPR. pp. 586–595 (2018) 9 24 Hyunjoon Park and Donghyeon Cho

    work page 2018

  43. [44]

    In: AAAI

    Zhang, W., Zhang, L., Hu, P., Ma, L., Zhuge, Y., Lu, H.: Bootstraping clustering of gaussians for view-consistent 3d scene understanding. In: AAAI. vol. 39, pp. 10166–10175 (2025) 3

    work page 2025

  44. [45]

    In: CVPR

    Zhou,S.,Chang,H.,Jiang,S.,Fan,Z.,Zhu,Z.,Xu,D.,Chari,P.,You,S.,Wang,Z., Kadambi, A.: Feature 3dgs: Supercharging 3d gaussian splatting to enable distilled feature fields. In: CVPR. pp. 21676–21685 (2024) 3, 10, 11

    work page 2024

  45. [46]

    IJCV133(2), 611–627 (2025) 3

    Zuo, X., Samangouei, P., Zhou, Y., Di, Y., Li, M.: Fmgs: Foundation model em- bedded 3d gaussian splatting for holistic 3d scene understanding. IJCV133(2), 611–627 (2025) 3

    work page 2025