Pith. sign in

REVIEW 5 major objections 5 minor 67 references

This paper claims that adding Gaussian-splatting attributes—opacity and scale—to 3D point clouds via Mahalanobis-matched correspondences makes semantic segmentation more accurate on geometrically ambiguous objects like refrigerators and sho

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

2026-08-03 12:16 UTC pith:DVIO2PG4

load-bearing objection Plausible new GS-to-point attribute transfer for point-cloud segmentation; the alignment step is under-specified and the 'superior performance' claim is slightly overbroad, but the paper is worth refereeing. the 5 major comments →

arxiv 2601.03510 v3 pith:DVIO2PG4 submitted 2026-01-07 cs.CV

G2P: Gaussian-to-Point Attribute Alignment for Boundary-Aware 3D Segmentation

Hojun Song , Chae-yeong Song , Jeong-hun Hong , Chaewon Moon , Soo Ye Kim , Yiyi Liao , Jaehyup Lee , Sang-hyo Park This is my paper
classification cs.CV
keywords 3D semantic segmentationGaussian Splattingpoint cloudappearance transferboundary-awareMahalanobis matchingfeature distillation
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.

G2P tries to show that sparse, geometry-only point clouds can be enriched with appearance cues by transferring attributes from optimized 3D Gaussian splats. It establishes point-wise correspondences using Mahalanobis distance, which respects the anisotropic shape of each splat, then augments every point with a weighted opacity and scale. Opacity supplies view-consistent appearance information; scale, which is small at object boundaries, is exploited to generate boundary pseudo-labels. On indoor benchmarks, this raises mIoU scores and yields the largest class-level gains on objects that are coplanar with walls or reflective. The approach requires no 2D image features or language pretraining in its segmentation pipeline.

Core claim

The central claim is that the misalignment between optimized Gaussians and original point geometry can be overcome by distance-based matching, and that the transferred attributes—opacity as appearance, scale as boundary cue—carry complementary information that resolves geometric bias. Reports 78.4 mIoU on ScanNet v2 (up 1.4 points over its reproduced baseline), 36.6 mIoU on ScanNet200 (up 2.1), and the biggest class improvements on refrigerator (+6.0 IoU) and shower curtain (+7.2). Qualitatively, the method fixes category confusion (e.g., door→wall) and incomplete coverage of thin coplanar structures.

What carries the argument

Gaussian-to-Point feature augmentation: for each point, Gaussian neighbors are collected within a radius and ranked by Mahalanobis distance (using each splat's covariance matrix). Inverse-distance weights aggregate the scale and opacity, producing a 13-dimensional augmented point. This feeds a self-supervised appearance encoder trained from scratch, which distills its features into the segmentation network via a cosine-similarity loss, while a scale-based extractor (points with small aggregate scale, after excluding background classes) supplies boundary pseudo-labels that supervise a boundary head.

Load-bearing premise

The load-bearing premise is that the optimized Gaussians stay in the same coordinate frame as the source point clouds and that a fixed Mahalanobis k-NN matching yields meaningful point-wise correspondences for every evaluation scene; the paper states both sets are aligned in the same coordinate space but gives no alignment procedure.

What would settle it

Re-run the pipeline on ScanNet v2 with Gaussians produced by a different optimizer or with their centroids randomly perturbed by a few centimeters; if the reported mIoU gains on refrigerator and shower curtain disappear, the alignment assumption is carrying the result.

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

If this is right

  • Segmentation of coplanar and reflective objects—doors, windows, refrigerators, shower curtains—improves substantially over geometry-only models.
  • The method transfers to a larger label space (ScanNet200) and to other indoor datasets, with consistent gains over the same backbone.
  • The appearance encoder is trained from scratch per dataset, so no external 2D or language supervision is needed in the pipeline.
  • Inference uses only point clouds; Gaussian attributes are consumed during training, so the deployed model is a standard point cloud segmenter.
  • Covariance-aware (Mahalanobis) matching with a moderate neighborhood size (k=20) outperforms Euclidean matching and larger or smaller neighborhoods.

Where Pith is reading between the lines

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

  • If small-scale-at-boundaries holds in cluttered scenes, the same scale-based pseudo-labels could be used for weakly-supervised or semi-supervised segmentation, where manual boundary labels are scarce.
  • The alignment step is stated but not specified; if Mahalanobis matching works across different Gaussian-splatting optimizers, G2P becomes a drop-in plug-in for any point cloud model, but if not, the gains may be tied to the particular reconstruction pipeline.
  • The boundary pseudo-labels are computed after excluding background classes; a testable variant is to derive them from scale alone or from opacity alone to isolate which attribute drives the boundary benefit.
  • Because distillation removes the need for Gaussians at inference, the same augmentation could be extended to point cloud instance segmentation or detection, where appearance cues also resolve geometric ambiguity.

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

5 major / 5 minor

Summary. The paper proposes G2P, a method that augments point cloud segmentation with attributes (opacity, scale) transferred from 3D Gaussian Splatting primitives. Correspondence is established via Mahalanobis-distance k-NN matching between points and Gaussians, and the aggregated attributes are used for two purposes: distilling appearance features from a pre-trained self-supervised encoder into a PT v3 backbone, and generating scale-based boundary pseudo-labels. The method is evaluated on ScanNet v2, ScanNet200, ScanNet++, and Matterport3D, reporting 78.4 mIoU on ScanNet v2 and 36.6 mIoU on ScanNet200, with ablations showing each component contributes about +0.8 mIoU.

Significance. If the core alignment premise holds, G2P is a genuinely useful idea: it transfers appearance cues in native 3D space without 2D or language supervision, and the ablations consistently attribute the gains to the proposed components. The paper is well-structured and includes multiple benchmarks and an efficiency analysis. However, the load-bearing assumption that SceneSplat-7K Gaussians are coordinate-aligned with the point clouds for every evaluation scene is asserted rather than demonstrated. The reported gains are small (+0.4 to +2.1 mIoU), leaving room for alignment artifacts to explain part of the improvement. The paper also overstates its 'superior performance' on the ScanNet200 hidden test, where it is 0.5 mIoU below BFANet. These issues require substantial revision, not just polishing.

major comments (5)
  1. [Sec. 3.3, 4.1, Eq. (1)-(2)] The alignment between SceneSplat-7K Gaussians and source point clouds is asserted but not demonstrated. Sec. 3.3 states 'Both sets are first aligned in the same coordinate space' with no procedure, and Sec. 4.1 only says 'we adopt the SceneSplat-7K dataset' without explaining how GS reconstructions are generated or aligned for ScanNet v2, ScanNet200, ScanNet++, Matterport3D, and the hidden test. If the Gaussians are not rigidly aligned, the Mahalanobis distances in Eq. (1) and the weighted aggregates in Eq. (2) become noise, invalidating both the appearance distillation and the scale-based boundary pseudo-labels. The candidate radius r in Sec. 3.3 is also never specified (the r=4 cm in Sec. 4.1 is for the semantic boundary in Sec. 3.4, a different quantity). Please provide the alignment procedure, the source of the GS data, and per-scene alignment verification (e.g., residual errors) for
  2. [Table 2, Sec. 4.2] The division into 'geometrically distinguishable' and 'geometrically challenging' classes is post hoc and not defined by any criterion. The challenging subset contains only 8 classes, and the claim of 'significant improvements' is not supported by statistical testing. In fact, the 'Picture' class drops from 41.8 (PT v3) to 39.4 (G2P), and the 'Shower Curtain' gain of +7.2 and 'Refrigerator' gain of +6.0 drive the subset average. A principled, pre-registered definition of geometric challenge (e.g., based on boundary ambiguity or appearance-geometry agreement) is needed, along with confidence intervals across multiple runs, before the class-level claim is credible.
  3. [Table 4, Abstract] The abstract claims 'superior performance on standard benchmarks' and 'significant improvements on geometrically challenging classes,' but on the ScanNet200 hidden test (Table 4) G2P achieves 35.5 mIoU, which is 0.5 below BFANet's 36.0. This undercuts the unconditional superiority claim. The paper should explicitly acknowledge this result and contextualize it (e.g., state that G2P matches BFANet on Head classes but trails on Common/Tail). As written, the claim is overstated relative to the reported evidence.
  4. [Tables 1, 5, 8] All results are single-run numbers with no error bars or significance tests. Differences of +0.4 to +1.4 mIoU are small enough that run-to-run variance could alter the ranking (e.g., Table 8 shows k=20 at 78.4 vs. k=30 at 77.9, a 0.5 gap that is presumably not significant). Given that the central gains are around 1 mIoU, the paper should report at least 3 seeds with mean and standard deviation, or provide significance testing, to establish that the improvements are real rather than noise.
  5. [Sec. 3.4, Fig. 3] The boundary pseudo-label generation relies on the assumption that small Gaussian scales concentrate at object boundaries. This is qualitatively illustrated in Fig. 3 but never quantitatively validated. The paper does not report the precision/recall of the resulting boundary pseudo-labels against ground-truth semantic edges, nor the sensitivity of the final mIoU to the trimming ratio τ_η (only a single value, 0.7, is used). Without such evidence, it is unclear whether the proposed scale-based extraction is actually detecting boundaries or merely encoding some correlated bias. Please add a boundary-quality evaluation and a sensitivity analysis for τ_η.
minor comments (5)
  1. [Eq. (5), Sec. 4.1] The notation is confusing: Eq. (5) defines a threshold τ_η as pruning the top η% of points, but Sec. 4.1 says 'we set the scale-based trimming ratio τ_η = 0.7.' Please clarify whether η=0.7 (keep 30% as boundary) or the threshold value is 0.7, and update the notation accordingly.
  2. [Sec. 3.5] The appearance encoder is said to follow Sonata, but the self-supervised pretext task (masked autoencoding, contrastive, etc.) is not described. Please specify the architecture and training objective so that the pre-training stage is reproducible.
  3. [Sec. 4.1] The source and coverage of SceneSplat-7K are not stated clearly. It is described as containing indoor scenes, but the paper does not say how many scene reconstructions are used per benchmark, whether the same GS reconstructions are used for ScanNet v2 and ScanNet200, or whether Matterport3D and ScanNet++ have their own GS sets. This is difficult to assess without a dataset table.
  4. [Abstract, Sec. 1] The phrase 'all without any 2D or language supervision' is slightly misleading because the semantic boundary pseudo-labels in Eq. (5)-(6) use ground-truth semantic labels to form B_sem, and the method borrows the B-S block from BFANet. The paper should explicitly distinguish 'no auxiliary 2D/language supervision' from 'no use of semantic labels beyond training supervision.'
  5. [Table 2] The formatting of Table 2 is hard to parse: the columns for the challenging classes are crowded, and the 'Avg.' columns are not clearly separated from the class columns. Please restructure the table for readability.

Circularity Check

0 steps flagged

No significant circularity: G2P's augmentation, boundary labels, and distillation are not defined in terms of the reported segmentation targets.

full rationale

I checked the derivation chain against the seven circularity patterns. The core equations are non-circular: Eqs. (1)-(2) aggregate external GS attributes (opacity and scale) by Mahalanobis distance, and no target mIoU or semantic label appears in these definitions. The boundary pseudo-labels of Eqs. (4)-(5) threshold aggregated GS scale magnitudes after filtering background classes with ground-truth semantics; B_sem also uses GT semantic boundaries, but this is ordinary auxiliary supervision for a boundary head rather than a prediction that reduces to the reported metric. The appearance teacher is pretrained from scratch with the self-supervised Sonata framework on (mu_p, c, alpha) without semantic labels, and Eq. (6) distills its features into the student; the teacher is not constructed from the student's outputs or from the benchmark scores. Hyperparameters (k, tau_eta, lambda) are tuned on validation, which is model selection rather than definitional circularity. No load-bearing self-citation chain exists: SceneSplat-7K [30], Sonata [53], PT v3 [52], and BFANet [63] are external works, and the paper does not invoke any author-specific uniqueness theorem. The main validity concern - the asserted but undocumented coordinate alignment between SceneSplat-7K Gaussians and point clouds in Sec. 3.3 - is a reproducibility/robustness risk, not a circularity: if the alignment failed, the pipeline would be undefined, but the method's equations would still not be equivalent to their inputs. Therefore no circular step is present.

Axiom & Free-Parameter Ledger

5 free parameters · 7 axioms · 0 invented entities

The central method is an empirical pipeline rather than a derivation. It leans on several domain assumptions about Gaussian Splatting attributes (opacity as appearance, scale as boundary indicator, aligned coordinate frames) and on external datasets/reconstructions; it introduces no new entities beyond an augmented point feature vector. The hyperparameters τ_η, k, r, λ_d, λ_b are tuned on validation and are the main fitted quantities.

free parameters (5)
  • Boundary pruning ratio τ_η = 0.7
    Top 70% of largest scale-magnitude points are pruned to define boundary pseudo-labels (Eq. 5). Chosen by hand/validation; directly controls B_scale.
  • Neighborhood size k for Mahalanobis matching = 20
    Selected after ablation (Table 8); k=20 gives best mIoU. Affects aggregated scale/opacity per point.
  • Semantic boundary radius r = 4 cm
    Radius used to detect neighboring points with different semantic labels for B_sem (Sec. 3.4).
  • Distillation loss weight λ_d = 0.4
    Weight for appearance distillation loss in Eq. 9.
  • Boundary loss weight λ_b = 0.9
    Weight for boundary loss in Eq. 9.
axioms (7)
  • standard math Mahalanobis distance, covariance inversion, and k-NN averaging (Eqs. 1-2) are valid mathematical operations.
    Used to compute point-Gaussian weights; no novel math claimed.
  • domain assumption Small Gaussian scales concentrate at object boundaries; large scales cover planar regions.
    Basis for scale-based boundary pseudo-labels (Eqs. 4-5); an empirical property of GS optimization that could fail in cluttered or fragmented geometry. Referenced in Fig. 3 and Sec. 3.4.
  • domain assumption Opacity α encodes appearance/visibility cues that are view-consistent and discriminative for semantic segmentation.
    Used to pretrain the appearance encoder on (μ,c,α) and for distillation (Sec. 3.5); supported by citations [30,34,62] but not demonstrated here in isolation.
  • domain assumption Optimized Gaussians and point clouds can be aligned in the same coordinate space without an explicit, described alignment procedure.
    Sec. 3.3 says 'Both sets are first aligned in the same coordinate space' but gives no method for obtaining this alignment for every test scene; if this alignment is unavailable or inaccurate, the augmentation cannot be applied.
  • domain assumption SceneSplat-7K reconstructions are available and aligned for ScanNet v2, ScanNet200, ScanNet++, and Matterport3D evaluation scenes.
    Sec. 4.1 says G2P adopts SceneSplat-7K; the paper does not state how GS for datasets beyond SceneSplat's coverage (e.g., Matterport3D) are produced or aligned.
  • ad hoc to paper Background classes (floor, wall) can be manually identified and removed before scale-based boundary extraction.
    Sec. 3.4 excludes 'background classes (e.g., floor, wall)' from P'_obj; this requires a dataset-specific class list and could bias boundary pseudo-labels.
  • domain assumption Sonata self-supervised pretraining on dataset-specific augmented points yields a teacher encoder whose features improve the downstream segmentation model via distillation.
    Sec. 3.5 uses Sonata architecture pretrained from scratch; the paper does not ablate the choice of pretraining objective.

pith-pipeline@v1.3.0-alltime-deepseek · 15223 in / 18660 out tokens · 160447 ms · 2026-08-03T12:16:24.053547+00:00 · methodology

0 comments
read the original abstract

Point cloud segmentation is critical for 3D scene understanding. However, sparse and irregular point distributions provide limited appearance evidence, making geometry-only features insufficient to distinguish objects with similar shapes but distinct appearances e.g., color, texture, and material. We propose Gaussian-to-Point (G2P), which transfers Gaussian attributes from 3D Gaussian Splatting to point clouds for more discriminative and appearance-consistent segmentation. Our G2P addresses the misalignment between optimized Gaussians and original point geometry by establishing point-wise correspondences. By distilling opacity-derived visibility cues, we mitigate the geometric ambiguity that limits existing models. Additionally, Gaussian scale attributes enable precise boundary localization in complex 3D scenes. Extensive experiments demonstrate that our approach achieves competitive performance on standard benchmarks and shows notable improvements on geometrically challenging classes, without pretrained 2D features or language supervision in our segmentation pipeline.

Figures

Figures reproduced from arXiv: 2601.03510 by Chaewon Moon, Chae-yeong Song, Hojun Song, Jaehyup Lee, Jeong-hun Hong, Sang-hyo Park, Soo Ye Kim, Yiyi Liao.

Figure 1
Figure 1. Figure 1: G2P augments Gaussian attributes to points while preserving geometry. (a) Point Cloud. Traditional point cloud representations lack sufficient appearance information to distin￾guish objects with similar geometry. As shown in the prediction, the existing model fails to segment coplanar windows cleanly and misclassifies the metallic refrigerator (yellow boxes), revealing ge￾ometric bias from insufficient app… view at source ↗
Figure 2
Figure 2. Figure 2: Overview of the Gaussian-to-Point (G2P) learning approach. G2P comprises a preparation stage and a main training stage. In the preparation stage, (a) Gaussian-to-Point Feature Augmentation aligns 3D Gaussian (µ g , Σ, α) and point (µ p , c, n), aggregating Gaussian attributes (α, S), where S represents the scale derived from the Σ, into augmented point features (α ′ , S′ ). (b) Scale-based Boundary Extract… view at source ↗
Figure 3
Figure 3. Figure 3: Gaussian Scale distribution. (a) Points after alignment, color-coded by aggregated Gaussian scale magnitude (magenta = small, yellow = large). Small scales concentrate at object bound￾aries. (b) Extracted boundary pseudo-labels by filtering out large￾scale points. tify its k most relevant Gaussian neighbors. First, to ef￾ficiently narrow down candidates, we select all Gaussians whose centroids fall within … view at source ↗
Figure 4
Figure 4. Figure 4 [PITH_FULL_IMAGE:figures/full_fig_p007_4.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 · 4 linked inside Pith

  1. [1]

    The lov´asz-softmax loss: A tractable surrogate for the optimization of the intersection-over-union measure in neural networks

    Maxim Berman, Amal Rannen Triki, and Matthew B Blaschko. The lov´asz-softmax loss: A tractable surrogate for the optimization of the intersection-over-union measure in neural networks. InProceedings of the IEEE conference on computer vision and pattern recognition, pages 4413–4421,

  2. [2]

    On a measure of divergence between two multinomial populations.Sankhy ¯a: the indian journal of statistics, pages 401–406, 1946

    Anil Bhattacharyya. On a measure of divergence between two multinomial populations.Sankhy ¯a: the indian journal of statistics, pages 401–406, 1946. 4

    1946

  3. [3]

    Segment any 3d gaussians

    Jiazhong Cen, Jiemin Fang, Chen Yang, Lingxi Xie, Xi- aopeng Zhang, Wei Shen, and Qi Tian. Segment any 3d gaussians. InProceedings of the AAAI Conference on Ar- tificial Intelligence, pages 1971–1979, 2025. 2

    1971

  4. [4]

    Matterport3d: Learning from rgb-d data in indoor environments.arXiv preprint arXiv:1709.06158, 2017

    Angel Chang, Angela Dai, Thomas Funkhouser, Maciej Halber, Matthias Niessner, Manolis Savva, Shuran Song, Andy Zeng, and Yinda Zhang. Matterport3d: Learning from rgb-d data in indoor environments.arXiv preprint arXiv:1709.06158, 2017. 6, 8

    Pith/arXiv arXiv 2017

  5. [5]

    Bridging the domain gap: Self-supervised 3d scene under- standing with foundation models.Advances in Neural Infor- mation Processing Systems, 36:79226–79239, 2023

    Zhimin Chen, Longlong Jing, Yingwei Li, and Bing Li. Bridging the domain gap: Self-supervised 3d scene under- standing with foundation models.Advances in Neural Infor- mation Processing Systems, 36:79226–79239, 2023. 2

    2023

  6. [6]

    A unified point-based framework for 3d segmentation

    Hung-Yueh Chiang, Yen-Liang Lin, Yueh-Cheng Liu, and Winston H Hsu. A unified point-based framework for 3d segmentation. In2019 International Conference on 3D Vi- sion (3DV), pages 155–163. IEEE, 2019. 1

    2019

  7. [7]

    4d spatio-temporal convnets: Minkowski convolutional neural networks

    Christopher Choy, JunYoung Gwak, and Silvio Savarese. 4d spatio-temporal convnets: Minkowski convolutional neural networks. InProceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 3075–3084,

  8. [8]

    3dmv: Joint 3d-multi- view prediction for 3d semantic scene segmentation

    Angela Dai and Matthias Nießner. 3dmv: Joint 3d-multi- view prediction for 3d semantic scene segmentation. In Proceedings of the European conference on computer vision (ECCV), pages 452–468, 2018. 2

    2018

  9. [9]

    Scannet: Richly-annotated 3d reconstructions of indoor scenes

    Angela Dai, Angel X Chang, Manolis Savva, Maciej Hal- ber, Thomas Funkhouser, and Matthias Nießner. Scannet: Richly-annotated 3d reconstructions of indoor scenes. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 5828–5839, 2017. 5, 6

    2017

  10. [10]

    Bundlefusion: Real-time globally consistent 3d reconstruction using on-the-fly surface reintegration.ACM Transactions on Graphics (ToG), 36(4): 1, 2017

    Angela Dai, Matthias Nießner, Michael Zollh ¨ofer, Shahram Izadi, and Christian Theobalt. Bundlefusion: Real-time globally consistent 3d reconstruction using on-the-fly surface reintegration.ACM Transactions on Graphics (ToG), 36(4): 1, 2017. 1

    2017

  11. [11]

    Bgpseg: Boundary-guided prim- itive instance segmentation of point clouds.IEEE Transac- tions on Image Processing, 2025

    Zheng Fang, Chuanqing Zhuang, Zhengda Lu, Yiqun Wang, Lupeng Liu, and Jun Xiao. Bgpseg: Boundary-guided prim- itive instance segmentation of point clouds.IEEE Transac- tions on Image Processing, 2025. 2

    2025

  12. [12]

    Learning 3d semantic segmentation with only 2d image supervision

    Kyle Genova, Xiaoqi Yin, Abhijit Kundu, Caroline Panto- faru, Forrester Cole, Avneesh Sud, Brian Brewington, Brian Shucker, and Thomas Funkhouser. Learning 3d semantic segmentation with only 2d image supervision. In2021 In- ternational Conference on 3D Vision (3DV), pages 361–372. IEEE, 2021. 2

    2021

  13. [13]

    Boundary-aware geometric en- coding for semantic segmentation of point clouds

    Jingyu Gong, Jiachen Xu, Xin Tan, Jie Zhou, Yanyun Qu, Yuan Xie, and Lizhuang Ma. Boundary-aware geometric en- coding for semantic segmentation of point clouds. InPro- ceedings of the AAAI Conference on Artificial Intelligence, pages 1424–1432, 2021. 2

    2021

  14. [14]

    Sugar: Surface- aligned gaussian splatting for efficient 3d mesh reconstruc- tion and high-quality mesh rendering

    Antoine Gu ´edon and Vincent Lepetit. Sugar: Surface- aligned gaussian splatting for efficient 3d mesh reconstruc- tion and high-quality mesh rendering. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5354–5363, 2024. 2, 4

    2024

  15. [15]

    Deep learning for 3d point clouds: A survey.IEEE transactions on pattern analysis and machine intelligence, 43(12):4338–4364, 2020

    Yulan Guo, Hanyun Wang, Qingyong Hu, Hao Liu, Li Liu, and Mohammed Bennamoun. Deep learning for 3d point clouds: A survey.IEEE transactions on pattern analysis and machine intelligence, 43(12):4338–4364, 2020. 1, 2

    2020

  16. [16]

    All in one: Visual-description-guided unified point cloud segmen- tation

    Zongyan Han, Mohamed El Amine Boudjoghra, Jiahua Dong, Jinhong Wang, and Rao Muhammad Anwer. All in one: Visual-description-guided unified point cloud segmen- tation. InProceedings of the IEEE/CVF International Con- ference on Computer Vision, pages 24835–24845, 2025. 1

    2025

  17. [17]

    Refersplat: Referring segmentation in 3d gaussian splatting.arXiv preprint arXiv:2508.08252, 2025

    Shuting He, Guangquan Jie, Changshuo Wang, Yun Zhou, Shuming Hu, Guanbin Li, and Henghui Ding. Refersplat: Referring segmentation in 3d gaussian splatting.arXiv preprint arXiv:2508.08252, 2025. 2

    Pith/arXiv arXiv 2025

  18. [18]

    Bidirectional projection network for cross dimension scene understanding

    Wenbo Hu, Hengshuang Zhao, Li Jiang, Jiaya Jia, and Tien-Tsin Wong. Bidirectional projection network for cross dimension scene understanding. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14373–14382, 2021. 2, 5

    2021

  19. [19]

    Jsenet: Joint semantic segmentation and edge detection network for 3d point clouds

    Zeyu Hu, Mingmin Zhen, Xuyang Bai, Hongbo Fu, and Chiew-lan Tai. Jsenet: Joint semantic segmentation and edge detection network for 3d point clouds. InEuropean confer- ence on computer vision, pages 222–239. Springer, 2020. 1, 2

    2020

  20. [20]

    Odin: A single model for 2d and 3d segmentation

    Ayush Jain, Pushkal Katara, Nikolaos Gkanatsios, Adam W Harley, Gabriel Sarch, Kriti Aggarwal, Vishrav Chaudhary, and Katerina Fragkiadaki. Odin: A single model for 2d and 3d segmentation. InProceedings of the IEEE/CVF Con- ference on Computer Vision and Pattern Recognition, pages 3564–3574, 2024. 1, 2, 5, 6

    2024

  21. [21]

    Identity-aware language gaussian splatting for open-vocabulary 3d semantic segmen- tation

    SungMin Jang and Wonjun Kim. Identity-aware language gaussian splatting for open-vocabulary 3d semantic segmen- tation. InProceedings of the IEEE/CVF International Con- ference on Computer Vision, pages 20467–20476, 2025. 2

    2025

  22. [22]

    Multi-view pointnet for 3d scene understanding

    Maximilian Jaritz, Jiayuan Gu, and Hao Su. Multi-view pointnet for 3d scene understanding. InProceedings of the IEEE/CVF international conference on computer vision workshops, pages 0–0, 2019. 1, 2

    2019

  23. [23]

    Kim Jun-Seong, GeonU Kim, Kim Yu-Ji, Yu-Chiang Frank Wang, Jaesung Choe, and Tae-Hyun Oh. Dr. splat: Directly referring 3d gaussian splatting via direct language embed- ding registration. InProceedings of the Computer Vision and Pattern Recognition Conference, pages 14137–14146, 2025. 2 9

    2025

  24. [24]

    3d gaussian splatting for real-time radiance field rendering.ACM Transactions on Graphics, 42 (4), 2023

    Bernhard Kerbl, Georgios Kopanas, Thomas Leimk ¨uhler, and George Drettakis. 3d gaussian splatting for real-time radiance field rendering.ACM Transactions on Graphics, 42 (4), 2023. 2, 4

    2023

  25. [25]

    Segment any- thing

    Alexander Kirillov, Eric Mintun, Nikhila Ravi, Hanzi Mao, Chloe Rolland, Laura Gustafson, Tete Xiao, Spencer White- head, Alexander C Berg, Wan-Yen Lo, et al. Segment any- thing. InProceedings of the IEEE/CVF International Con- ference on Computer Vision, pages 4015–4026, 2023. 3

    2023

  26. [26]

    Oneformer3d: One transformer for unified point cloud segmentation

    Maxim Kolodiazhnyi, Anna V orontsova, Anton Konushin, and Danila Rukhovich. Oneformer3d: One transformer for unified point cloud segmentation. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 20943–20953, 2024. 1, 6

    2024

  27. [27]

    Virtual multi-view fusion for 3d semantic segmentation

    Abhijit Kundu, Xiaoqi Yin, Alireza Fathi, David Ross, Brian Brewington, Thomas Funkhouser, and Caroline Pantofaru. Virtual multi-view fusion for 3d semantic segmentation. In European conference on computer vision, pages 518–535. Springer, 2020. 2, 5

    2020

  28. [28]

    Joint learning of 2d- 3d weakly supervised semantic segmentation

    Hyeokjun Kweon and Kuk-Jin Yoon. Joint learning of 2d- 3d weakly supervised semantic segmentation. InAdvances in Neural Information Processing Systems, pages 30499– 30511. Curran Associates, Inc., 2022. 2

    2022

  29. [29]

    Pointcnn: Convolution on x-transformed points.Advances in neural information processing systems, 31, 2018

    Yangyan Li, Rui Bu, Mingchao Sun, Wei Wu, Xinhan Di, and Baoquan Chen. Pointcnn: Convolution on x-transformed points.Advances in neural information processing systems, 31, 2018. 2

    2018

  30. [30]

    Scenesplat: Gaussian splatting-based scene understanding with vision-language pretraining

    Yue Li, Qi Ma, Runyi Yang, Huapeng Li, Mengjiao Ma, Bin Ren, Nikola Popovic, Nicu Sebe, Ender Konukoglu, Theo Gevers, et al. Scenesplat: Gaussian splatting-based scene understanding with vision-language pretraining. InProceed- ings of the IEEE/CVF International Conference on Com- puter Vision (ICCV), 2025. 2, 3, 6

    2025

  31. [31]

    Rea- songrounder: Lvlm-guided hierarchical feature splatting for open-vocabulary 3d visual grounding and reasoning

    Zhenyang Liu, Yikai Wang, Sixiao Zheng, Tongying Pan, Longfei Liang, Yanwei Fu, and Xiangyang Xue. Rea- songrounder: Lvlm-guided hierarchical feature splatting for open-vocabulary 3d visual grounding and reasoning. InPro- ceedings of the Computer Vision and Pattern Recognition Conference, pages 3718–3727, 2025. 2

    2025

  32. [32]

    Decoupled weight decay regularization.arXiv preprint arXiv:1711.05101, 2017

    Ilya Loshchilov and Frank Hutter. Decoupled weight decay regularization.arXiv preprint arXiv:1711.05101, 2017. 6

    Pith/arXiv arXiv 2017

  33. [33]

    Scaffold-gs: Structured 3d gaussians for view-adaptive rendering

    Tao Lu, Mulin Yu, Linning Xu, Yuanbo Xiangli, Limin Wang, Dahua Lin, and Bo Dai. Scaffold-gs: Structured 3d gaussians for view-adaptive rendering. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 20654–20664, 2024. 2, 4

    2024

  34. [34]

    A large-scale dataset of gaussian splats and their self-supervised pretrain- ing

    Qi Ma, Yue Li, Bin Ren, Nicu Sebe, Ender Konukoglu, Theo Gevers, Luc Van Gool, and Danda Pani Paudel. A large-scale dataset of gaussian splats and their self-supervised pretrain- ing. In2025 International Conference on 3D Vision (3DV), pages 145–155. IEEE, 2025. 3

    2025

  35. [35]

    V-net: Fully convolutional neural networks for volumetric medical image segmentation

    Fausto Milletari, Nassir Navab, and Seyed-Ahmad Ahmadi. V-net: Fully convolutional neural networks for volumetric medical image segmentation. In2016 fourth international conference on 3D vision (3DV), pages 565–571. Ieee, 2016. 5

    2016

  36. [36]

    Pointnet: Deep learning on point sets for 3d classification and segmentation

    Charles R Qi, Hao Su, Kaichun Mo, and Leonidas J Guibas. Pointnet: Deep learning on point sets for 3d classification and segmentation. InProceedings of the IEEE conference on computer vision and pattern recognition, pages 652–660,

  37. [37]

    Pointnet++: Deep hierarchical feature learning on point sets in a metric space

    Charles Ruizhongtai Qi, Li Yi, Hao Su, and Leonidas J Guibas. Pointnet++: Deep hierarchical feature learning on point sets in a metric space. InAdvances in Neural Infor- mation Processing Systems. Curran Associates, Inc., 2017. 2

    2017

  38. [38]

    Langsplat: 3d language gaussian splatting

    Minghan Qin, Wanhua Li, Jiawei Zhou, Haoqian Wang, and Hanspeter Pfister. Langsplat: 3d language gaussian splatting. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 20051–20060, 2024. 3

    2024

  39. [39]

    Learning transferable visual models from natural language supervi- sion

    Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervi- sion. InInternational conference on machine learning, pages 8748–8763. PmLR, 2021. 3

    2021

  40. [40]

    Learn- ing multi-view aggregation in the wild for large-scale 3d se- mantic segmentation

    Damien Robert, Bruno Vallet, and Loic Landrieu. Learn- ing multi-view aggregation in the wild for large-scale 3d se- mantic segmentation. InProceedings of the IEEE/CVF Con- ference on Computer Vision and Pattern Recognition, pages 5575–5584, 2022. 1, 2, 5

    2022

  41. [41]

    Edge-aware 3d instance segmentation network with intelligent semantic prior

    Wonseok Roh, Hwanhee Jung, Giljoo Nam, Jinseop Yeom, Hyunje Park, Sang Ho Yoon, and Sangpil Kim. Edge-aware 3d instance segmentation network with intelligent semantic prior. InProceedings of the IEEE/CVF Conference on Com- puter Vision and Pattern Recognition, pages 20644–20653,

  42. [42]

    Language- grounded indoor 3d semantic segmentation in the wild

    David Rozenberszki, Or Litany, and Angela Dai. Language- grounded indoor 3d semantic segmentation in the wild. In European conference on computer vision, pages 125–141. Springer, 2022. 6

    2022

  43. [43]

    Indoorgs: Geometric cues guided gaussian splatting for indoor scene reconstruction

    Cong Ruan, Yuesong Wang, Tao Guan, Bin Zhang, and Lili Ju. Indoorgs: Geometric cues guided gaussian splatting for indoor scene reconstruction. InProceedings of the Computer Vision and Pattern Recognition Conference, pages 844–853,

  44. [44]

    Trace3d: Consistent segmen- tation lifting via gaussian instance tracing

    Hongyu Shen, Junfeng Ni, Yixin Chen, Weishuo Li, Ming- tao Pei, and Siyuan Huang. Trace3d: Consistent segmen- tation lifting via gaussian instance tracing. InProceedings of the IEEE/CVF International Conference on Computer Vi- sion, pages 6656–6666, 2025. 3

    2025

  45. [45]

    Flashsplat: 2d to 3d gaussian splatting segmentation solved optimally

    Qiuhong Shen, Xingyi Yang, and Xinchao Wang. Flashsplat: 2d to 3d gaussian splatting segmentation solved optimally. In European Conference on Computer Vision, pages 456–472. Springer, 2024. 2

    2024

  46. [46]

    Multi-view convolutional neural networks for 3d shape recognition

    Hang Su, Subhransu Maji, Evangelos Kalogerakis, and Erik Learned-Miller. Multi-view convolutional neural networks for 3d shape recognition. InProceedings of the IEEE in- ternational conference on computer vision, pages 945–953,

  47. [47]

    Contrastive boundary learning for point cloud segmentation

    Liyao Tang, Yibing Zhan, Zhe Chen, Baosheng Yu, and Dacheng Tao. Contrastive boundary learning for point cloud segmentation. InProceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 8489– 8499, 2022. 1, 2 10

    2022

  48. [48]

    Octformer: Octree-based transformers for 3d point clouds.ACM Transactions on Graphics (TOG), 42(4):1–11, 2023

    Peng-Shuai Wang. Octformer: Octree-based transformers for 3d point clouds.ACM Transactions on Graphics (TOG), 42(4):1–11, 2023. 1, 2, 5, 6, 7, 8

    2023

  49. [49]

    Unipre3d: Unified pre-training of 3d point cloud models with cross-modal gaussian splatting

    Ziyi Wang, Yanran Zhang, Jie Zhou, and Jiwen Lu. Unipre3d: Unified pre-training of 3d point cloud models with cross-modal gaussian splatting. InProceedings of the Computer Vision and Pattern Recognition Conference, pages 1319–1329, 2025. 3, 5, 6

    2025

  50. [50]

    Dc- seg: Decoupled 3d open-set segmentation using gaussian splatting

    Luis Wiedmann, Luca Wiehe, and David Rozenberszki. Dc- seg: Decoupled 3d open-set segmentation using gaussian splatting. InProceedings of the Computer Vision and Pat- tern Recognition Conference, pages 5217–5226, 2025. 2

    2025

  51. [51]

    Pointconv: Deep convolutional networks on 3d point clouds

    Wenxuan Wu, Zhongang Qi, and Li Fuxin. Pointconv: Deep convolutional networks on 3d point clouds. InProceedings of the IEEE/CVF Conference on computer vision and pattern recognition, pages 9621–9630, 2019. 2

    2019

  52. [52]

    Point transformer v3: Simpler faster stronger

    Xiaoyang Wu, Li Jiang, Peng-Shuai Wang, Zhijian Liu, Xi- hui Liu, Yu Qiao, Wanli Ouyang, Tong He, and Hengshuang Zhao. Point transformer v3: Simpler faster stronger. InPro- ceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 4840–4851, 2024. 1, 2, 5, 6, 7, 8

    2024

  53. [53]

    Sonata: Self- supervised learning of reliable point representations

    Xiaoyang Wu, Daniel DeTone, Duncan Frost, Tianwei Shen, Chris Xie, Nan Yang, Jakob Engel, Richard New- combe, Hengshuang Zhao, and Julian Straub. Sonata: Self- supervised learning of reliable point representations. InPro- ceedings of the Computer Vision and Pattern Recognition Conference, pages 22193–22204, 2025. 2, 5, 6

    2025

  54. [54]

    Opengaussian: Towards point-level 3d gaussian-based open vocabulary understanding.Advances in Neural Information Processing Systems, 37:19114–19138,

    Yanmin Wu, Jiarui Meng, Haijie Li, Chenming Wu, Yahao Shi, Xinhua Cheng, Chen Zhao, Haocheng Feng, Errui Ding, Jingdong Wang, et al. Opengaussian: Towards point-level 3d gaussian-based open vocabulary understanding.Advances in Neural Information Processing Systems, 37:19114–19138,

  55. [55]

    Pointcontrast: Unsupervised pre- training for 3d point cloud understanding

    Saining Xie, Jiatao Gu, Demi Guo, Charles R Qi, Leonidas Guibas, and Or Litany. Pointcontrast: Unsupervised pre- training for 3d point cloud understanding. InEuropean con- ference on computer vision, pages 574–591. Springer, 2020. 2, 6

    2020

  56. [56]

    In- vestigate indistinguishable points in semantic segmentation of 3d point cloud

    Mingye Xu, Zhipeng Zhou, Junhao Zhang, and Yu Qiao. In- vestigate indistinguishable points in semantic segmentation of 3d point cloud. InProceedings of the AAAI Conference on Artificial Intelligence, pages 3047–3055, 2021. 2

    2021

  57. [57]

    2d-3d interlaced transformer for point cloud segmentation with scene-level supervision

    Cheng-Kun Yang, Min-Hung Chen, Yung-Yu Chuang, and Yen-Yu Lin. 2d-3d interlaced transformer for point cloud segmentation with scene-level supervision. InProceedings of the IEEE/CVF International Conference on Computer Vi- sion, pages 977–987, 2023. 2

    2023

  58. [58]

    Gaussian grouping: Segment and edit anything in 3d scenes

    Mingqiao Ye, Martin Danelljan, Fisher Yu, and Lei Ke. Gaussian grouping: Segment and edit anything in 3d scenes. InEuropean conference on computer vision, pages 162–179. Springer, 2024. 2

    2024

  59. [59]

    Scannet++: A high-fidelity dataset of 3d in- door scenes

    Chandan Yeshwanth, Yueh-Cheng Liu, Matthias Nießner, and Angela Dai. Scannet++: A high-fidelity dataset of 3d in- door scenes. InProceedings of the IEEE/CVF International Conference on Computer Vision, pages 12–22, 2023. 5, 6, 8

    2023

  60. [60]

    Gaussian opacity fields: Efficient adaptive surface reconstruction in unbounded scenes.ACM Transactions on Graphics (ToG), 43(6):1–13, 2024

    Zehao Yu, Torsten Sattler, and Andreas Geiger. Gaussian opacity fields: Efficient adaptive surface reconstruction in unbounded scenes.ACM Transactions on Graphics (ToG), 43(6):1–13, 2024. 2

    2024

  61. [61]

    Panogs: Gaussian-based panoptic seg- mentation for 3d open vocabulary scene understanding

    Hongjia Zhai, Hai Li, Zhenzhe Li, Xiaokun Pan, Yijia He, and Guofeng Zhang. Panogs: Gaussian-based panoptic seg- mentation for 3d open vocabulary scene understanding. In Proceedings of the Computer Vision and Pattern Recognition Conference, pages 14114–14124, 2025. 2

    2025

  62. [62]

    Mitigating ambiguities in 3d classification with gaussian splatting

    Ruiqi Zhang, Hao Zhu, Jingyi Zhao, Qi Zhang, Xun Cao, and Zhan Ma. Mitigating ambiguities in 3d classification with gaussian splatting. InProceedings of the Computer Vision and Pattern Recognition Conference, pages 27275–27284,

  63. [63]

    Bfanet: Revisiting 3d semantic segmentation with boundary feature analysis

    Weiguang Zhao, Rui Zhang, Qiufeng Wang, Guangliang Cheng, and Kaizhu Huang. Bfanet: Revisiting 3d semantic segmentation with boundary feature analysis. InProceedings of the Computer Vision and Pattern Recognition Conference, pages 29395–29405, 2025. 1, 2, 5, 6, 7

    2025

  64. [64]

    Understanding imbalanced semantic segmentation through neural collapse

    Zhisheng Zhong, Jiequan Cui, Yibo Yang, Xiaoyang Wu, Xi- aojuan Qi, Xiangyu Zhang, and Jiaya Jia. Understanding imbalanced semantic segmentation through neural collapse. InProceedings of the IEEE/CVF conference on computer vi- sion and pattern recognition, pages 19550–19560, 2023. 7

    2023

  65. [65]

    Feature 3dgs: Supercharging 3d gaussian splatting to enable distilled feature fields

    Shijie Zhou, Haoran Chang, Sicheng Jiang, Zhiwen Fan, Ze- hao Zhu, Dejia Xu, Pradyumna Chari, Suya You, Zhangyang Wang, and Achuta Kadambi. Feature 3dgs: Supercharging 3d gaussian splatting to enable distilled feature fields. InPro- ceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 21676–21685, 2024. 3

    2024

  66. [66]

    Ponderv2: Pave the way for 3d foun- dation model with a universal pre-training paradigm.arXiv preprint arXiv:2310.08586, 2023

    Haoyi Zhu, Honghui Yang, Xiaoyang Wu, Di Huang, Sha Zhang, Xianglong He, Hengshuang Zhao, Chunhua Shen, Yu Qiao, Tong He, et al. Ponderv2: Pave the way for 3d foun- dation model with a universal pre-training paradigm.arXiv preprint arXiv:2310.08586, 2023. 5, 6, 7

    Pith/arXiv arXiv 2023

  67. [67]

    Rethinking end- to-end 2d to 3d scene segmentation in gaussian splatting

    Runsong Zhu, Shi Qiu, Zhengzhe Liu, Ka-Hei Hui, Qianyi Wu, Pheng-Ann Heng, and Chi-Wing Fu. Rethinking end- to-end 2d to 3d scene segmentation in gaussian splatting. In Proceedings of the Computer Vision and Pattern Recognition Conference, pages 3656–3665, 2025. 3 11

    2025