Pith. sign in

REVIEW 4 major objections 7 minor 1 cited by

VoteSplat: Hough Voting Gaussian Splatting for 3D Scene Understanding

T0 review · 4 major / 7 minor · reviewed 2026-08-06 · deepseek-v4-flash

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

Pith's one-line read VoteSplat turns 3D scene understanding into a voting problem, replacing per-Gaussian feature embeddings with offset vectors that point each Gaussian toward its instance center.

desk verdict VoteSplat is a promising integration of Hough voting into 3DGS with a genuinely cheaper training recipe, but the central voting transmittance in Eq. (7) is never defined, and that is a load-bearing gap, not a cosmetic one. read the letter →

arxiv 2506.22799 v1 pith:F3VS5YJP submitted 2025-06-28 cs.GR cs.CVcs.LG

Minchao Jiang , Shunyu Jia , Jiaming Gu , Xiaoyuan Lu , Guangming Zhu , Anqi Dong , Liang Zhang This is my paper
classification cs.GRcs.CVcs.LG
keywords 3DGaussianSplattingHoughvotinginstancesegmentationopen-vocabularylocalizationdepthregularizationsemanticassociationpointcloudunderstanding
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

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

The reading

VoteSplat tries to establish that 3D Gaussian Splatting scenes can be understood at the instance level without storing high-dimensional semantic features on every Gaussian. Its central idea is to give each Gaussian a learned 3D offset vector; adding that offset to the Gaussian's position produces a 'vote' that lands near the center of the object instance the Gaussian belongs to. These votes are trained against 2D centroids derived from SAM masks, then clustered to assign instance IDs. If this works, open-vocabulary object localization, click-based selection, and scene editing all follow from simple clustering of votes instead of expensive contrastive feature learning, which would make 3D scene understanding much cheaper to train.

What carries the argument

The central mechanism is 3D vote splatting: each Gaussian's offset vector $\Delta p_i$ is added to its center $p_i$ to form a 3D vote $V^{3d}_i = \Delta p_i + p_i$, rendered to screen space through a modified blending rule $V^{3d} = \frac{1}{|M|}\sum_{i\in M} V^{3d}_i$ that uses a voting transmittance set $M$ to exclude occluded and background Gaussians. The vote loss compares these projected votes to 2D mask centroids, and the depth distortion loss $L^*_d = \sum_{i,j}|z_i - z_j|$ ties votes together along depth. Clustering the resulting 3D votes with HDBSCAN yields instance IDs, which are then linked to one CLIP feature per instance.

What would settle it

Train VoteSplat on a scene with a heavily occluded object and then measure whether the 3D votes for that object's visible and occluded parts still converge within one tight cluster; if they scatter or merge with the occluder's votes, the central claim of instance-center voting fails.

Watch

Extended reading notes

Core claim

VoteSplat claims that instance-level 3D understanding of a Gaussian splat can be achieved by learning one extra 3D vector per Gaussian primitive: the spatial offset from the primitive to its instance centroid. During training, SAM produces 2D masks whose pixel centroids serve as ground-truth votes; the rendered projection of the 3D votes is supervised to match these 2D votes. A depth distortion regularizer pushes voting points together along the depth axis. After training, clustering the 3D votes with HDBSCAN separates instances, and each instance is matched to a single CLIP feature derived from its rendered pixels, enabling open-vocabulary queries and click-based editing without per-Gaussian language features.

Load-bearing premise

The method assumes that the set of Gaussians that should contribute to a given pixel's vote (excluding background and occluded ones) can be cleanly identified during training, but the paper does not specify how that set is computed.

Editorial extensions

If this is right

  • If VoteSplat works as claimed, 3D instance segmentation no longer requires per-Gaussian high-dimensional features; a single 3D offset per Gaussian can replace them, cutting training time relative to contrastive feature methods.
  • Open-vocabulary object localization becomes a one-step retrieval: the text query matches one CLIP vector per instance, not per pixel or per Gaussian.
  • Click-based 3D object selection and scene editing follow directly from the instance ID map, because selecting a pixel gives the whole object's Gaussian cluster.
  • Hierarchical SAM masks can be reused to supervise multi-level 3D votes, yielding part-level segmentation without a separate part-modeling stage.
  • Background Gaussians, which have zero offset, are filtered out before clustering, so only object-surfacing primitives contribute to instance votes.

Reading between the lines

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

  • Because the method derives its 2D ground-truth votes from SAM mask centroids, VoteSplat inherits SAM's failure modes: if SAM merges two touching objects into one mask, the learned votes will also merge them, and no later step can separate them.
  • The voting transmittance set $M$ is the hidden control knob; a testable extension would be to let the network learn $M$ instead of leaving it implicit, which could improve robustness on scenes with heavy occlusion.
  • The method should transfer to any segmentation foundation model that produces masks, since SAM is used only to generate 2D vote supervision; swapping in a stronger mask model would likely improve instance-boundary accuracy.
  • The depth distortion loss may over-constrain thin or elongated objects, potentially pulling votes from different parts of an object along the depth axis; this could be tested on scenes with long tabletop objects.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

4 major / 7 minor

Summary. The paper proposes VoteSplat, a 3D Gaussian Splatting (3DGS) framework for 3D scene understanding. Each Gaussian primitive is augmented with a 3D offset vector, and these offsets are trained so that, when splatted, they produce 3D votes near instance centers. Supervision comes from SAM-generated 2D instance masks: the mask centroids are used as 2D ground-truth votes, and the projected 3D votes are compared against them. A depth distortion regularization term is added to improve depth consistency, and a voting transmittance model is introduced to exclude background and occluded Gaussians during vote construction. At inference, zero-offset Gaussians are removed as background, remaining Gaussians are clustered with HDBSCAN into instance IDs, and CLIP features are associated with instance IDs via rendered instance-ID maps. Experiments on LeRF and 3D-OVS report improved open-vocabulary segmentation mIoU/mAcc and lower training time than LangSplat and OpenGaussian, along with qualitative results for click-based selection, editing, and hierarchical segmentation.

Significance. If the method works as described, the contribution is significant: it offers point-level 3D instance segmentation and open-vocabulary localization without storing high-dimensional CLIP features on every Gaussian, and it does so with substantially lower training cost than contrastive-learning alternatives. The training signal comes from external SAM masks and evaluation uses external CLIP queries and manual annotations, so the results are not circular in the definitional sense. The paper also demonstrates a useful practical application, click-based 3D object selection and scene editing. The main risk is not the overall idea but the underspecification of the central voting mechanism: Eq. (7) relies on an undefined voting transmittance and set M, and the ablation in Table 3 shows that this choice is responsible for a large performance gap. A clear, reproducible definition of M is necessary before the reported numbers can be fully trusted.

major comments (4)
  1. [Section 3.3, Eq. (7)] The voting transmittance \hat T_i and the set M of 'depth-ordered points' are never defined. Eq. (7) is the central operation that produces the blended 3D vote, and Section 3.4 later removes Gaussians with \Delta p = 0 as background. Whether a Gaussian receives vote-loss gradients depends entirely on membership in M, yet the paper gives no rule for constructing M. Table 3 shows that replacing the voting transmittance with the standard transmittance T drops mIoU from 85.66 to 58.72, so this is not a cosmetic omission. Please define \hat T_i and M precisely, state how they are computed during training (e.g., from depth thresholds, opacity, SAM masks, or connected components), and explain how occluded and background Gaussians are excluded before the loss is evaluated.
  2. [Section 3.3, Eqs. (10)-(11)] The depth distortion term is changed from \sum \omega_i \omega_j |z_i - z_j| to \sum |z_i - z_j| by dropping weights, with only a brief statement that votes should be equally concentrated in depth. The summation domain is not specified; as written, the sum runs over all pairs of Gaussians, which is O(N^2) and not obviously compatible with the tile-based rasterizer. More importantly, no formal argument or experiment justifies that removing the weights preserves the intended clustering pressure without distorting the reconstruction objective. Please specify the set of pairs over which the sum is taken, provide a derivation or empirical justification for dropping the weights, and report the effect of the weighted version on reconstruction quality and segmentation accuracy.
  3. [Section 3.4, Background Filtering] The statement that Gaussians with \Delta p = 0 are removed as background because they 'receive no gradient updates' creates a circular dependency. These Gaussians receive no vote-loss gradient only if Eq. (7) already excludes them through the set M, but M is undefined. If M includes background Gaussians during training, their offset vectors will be pulled toward foreground instance centers, and the \Delta p = 0 filter at inference will fail. Please clarify how training and inference are made consistent, and give the exact criterion that keeps background offsets at zero.
  4. [Table 3, Ablation Study] The quantitative ablation on the snacks scene reports only three configurations: voting transmittance without L_d (76.04 mIoU), standard transmittance with L_d (58.72 mIoU), and voting transmittance with L_d (85.66 mIoU). It omits the standard-transmittance-without-L_d baseline, so the individual contributions of the voting transmittance and the depth loss are not fully identifiable. The table also reports a single scene with no variance or significance information. Please report the full 2x2 ablation and, ideally, results across multiple scenes with standard deviations.
minor comments (7)
  1. [Notation, Footnote 1] The footnote states that |\cdot| denotes the L1 norm, but in Eq. (7) |M| denotes cardinality and in Eq. (9) |P| denotes cardinality. Please use distinct notation, e.g., \mathop{\mathrm{card}}(M) or |M|_0, to avoid ambiguity.
  2. [Section 3.2, Eq. (3)] The centroid formula uses round() on the continuous centroid. Please explain whether rounding is necessary for the vote map representation and whether it introduces sub-pixel localization error that affects the loss.
  3. [Section 4.1, Experimental Setup] The text says 'all datasets annotated by LangSplat.' Please clarify what this means: are the evaluation ground-truth masks manually annotated, or are they generated by LangSplat? The fairness of the comparison depends on this detail.
  4. [Section 4.1, Table 2] Training times are reported with a '~' prefix and no variance. Please state the exact protocol, including the number of iterations, hardware, and whether timing includes SAM preprocessing and HDBSCAN clustering.
  5. [Section 3.4, Clustering] The HDBSCAN hyperparameters (min_cluster_size, min_samples) and the CLIP similarity threshold used for selecting instances are not reported. These parameters are necessary for reproducibility.
  6. [Section 5, Concluding Remarks] The limitations section mentions failures for instances exceeding the field of view and for instances in concave containers, but these are not quantified. Adding failure-case examples or a small quantitative study would strengthen the paper.
  7. [References, Section 3.1] The recap of 3DGS cites reference [39] (Mip-Splatting) for the rendering equations, but Eq. (1) and Eq. (2) are standard 3DGS equations from reference [10]. Please align the citations.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: VoteSplat's predictions are supervised by external SAM masks and evaluated against independent benchmarks, not by its own fitted outputs.

full rationale

VoteSplat's derivation chain is not circular. The 2D vote supervision (Eqs. 3-4 and 9) is computed from SAM masks, an external segmenter, and the 3D offset vectors are optimized against those projected 2D vote targets; this is a fitted mapping from external inputs, not a prediction of the same quantity used to define the loss. The open-vocabulary selection in Section 4.1 uses CLIP text/image features and evaluates against independently annotated masks on 3D-OVS/LeRF-OVS, so the evaluation is not produced by the method's own training targets. The instance IDs are obtained by HDBSCAN clustering of trained votes, and CLIP features are extracted afterward from the associated pixels; this is a post-hoc association rather than a self-referential definition. There are no load-bearing self-citations: the paper builds on 3DGS [10], SAM [12], LangSplat [24], OpenGaussian [35], and 2DGS [9] as external prior work, none by the present authors. The most legitimate concern is the underspecified 'voting transmittance' \hat{T}_i and set M in Eq. (7), together with the Section 3.4 rule that removes Gaussians with Δp=0 as background; because the paper never defines how M is constructed or how background Gaussians are excluded during training, the central training signal is difficult to reproduce from the text. Underspecification is a correctness and reproducibility flaw, not a circularity: the paper does not define M in terms of the learned offsets, and the 3D votes are not equal by construction to the SAM mask centroids. The ablation in Table 3 shows that removing components degrades performance, which is consistent with an independent, though incompletely specified, derivation.

Assumptions & free parameters 5 free parameters · 7 assumptions · 2 invented entities

The central result rests on SAM mask quality, the existence of a well-defined voting transmittance set, HDBSCAN separating votes, and CLIP features being discriminative at the instance level. These are domain assumptions rather than derived theorems; the paper does not quantify their failure modes.

free parameters (5)
  • lambda_vote = not reported
    Weight of the vote loss in Eq. (12); likely tuned per scene, but no value or sensitivity analysis is given.
  • lambda_Ld = not reported
    Weight of the depth regularization term in Eq. (12); not reported in the text.
  • HDBSCAN min_cluster_size and min_samples = not reported
    These clustering parameters directly control the number and quality of instance clusters in Section 3.4, yet no values are given.
  • SAM automatic mask generation thresholds = not reported
    SAM parameters determine which masks and centroids become the 2D supervision in Section 3.2.
  • CLIP similarity selection threshold = not reported
    For open-vocabulary selection, the paper does not specify how many instances are selected per query or what similarity threshold is used.
assumptions (7)
  • standard math The 3DGS differentiable rasterizer and alpha-blending equations from [10] are correct and inherited unchanged.
    The method builds on Eqs. (1) and (2); any error in the rasterizer would propagate to the vote rendering.
  • domain assumption SAM-based masks provide accurate instance boundaries and stable 2D centroids across views.
    Section 3.2 uses SAM automatic masks as the only supervision for 3D votes; inconsistent masks would produce inconsistent votes.
  • domain assumption Each instance has a well-defined 3D centroid that projects onto the SAM mask centroid in most views.
    The 2D vote supervises the projection of the 3D vote; if the object center is not visible or is occluded, the target becomes unreliable.
  • ad hoc to paper The set M of depth-ordered points that participate in uniform averaging in Eq. (7) can be identified without knowing the instance assignment.
    Eq. (7) relies on a 'voting transmittance' that selects participating Gaussians, but the paper never defines this set.
  • ad hoc to paper Removing weights from the depth distortion loss in Eq. (11) still produces meaningful clustering pressure without harming reconstruction.
    The paper drops the alpha and transmittance weights from Eq. (10) to get Eq. (11) with no formal argument beyond 'the weight should not be reduced'.
  • domain assumption HDBSCAN on 3D votes recovers one cluster per object instance in complex scenes.
    Section 3.4 applies HDBSCAN directly; the paper shows qualitative separation but no robustness analysis across scenes.
  • domain assumption CLIP features of rendered instance pixels are semantically aligned with text queries.
    Open-vocabulary selection computes cosine similarity between per-instance CLIP features and CLIP text features; this assumes CLIP generalizes to rendered 3D crops.
invented entities (2)
  • Per-Gaussian 3D offset vector Δp_i
    purpose: Votes each Gaussian toward the center of its instance; the 3D vote is p_i + Δp_i.
    A model-internal parameter; its only support is the paper's own experiments, with no external falsifiable prediction.
  • Voting transmittance model
    purpose: Defines which depth-ordered Gaussians are averaged uniformly in Eq. (7) to exclude occluded and background points.
    Never defined formally; it cannot be checked or replicated from the text.

how reviews work

0 comments
Cite this review

Pith. "Pith review of VoteSplat: Hough Voting Gaussian Splatting for 3D Scene Understanding." pith.science (2026). https://pith.science/paper/F3VS5YJP

@misc{pith2026250622799,
  author       = {Pith},
  title        = {Pith review of: VoteSplat: Hough Voting Gaussian Splatting for 3D Scene Understanding},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/F3VS5YJP}},
  note         = {Machine review of arXiv:2506.22799}
}
read the original abstract

3D Gaussian Splatting (3DGS) has become horsepower in high-quality, real-time rendering for novel view synthesis of 3D scenes. However, existing methods focus primarily on geometric and appearance modeling, lacking deeper scene understanding while also incurring high training costs that complicate the originally streamlined differentiable rendering pipeline. To this end, we propose VoteSplat, a novel 3D scene understanding framework that integrates Hough voting with 3DGS. Specifically, Segment Anything Model (SAM) is utilized for instance segmentation, extracting objects, and generating 2D vote maps. We then embed spatial offset vectors into Gaussian primitives. These offsets construct 3D spatial votes by associating them with 2D image votes, while depth distortion constraints refine localization along the depth axis. For open-vocabulary object localization, VoteSplat maps 2D image semantics to 3D point clouds via voting points, reducing training costs associated with high-dimensional CLIP features while preserving semantic unambiguity. Extensive experiments demonstrate effectiveness of VoteSplat in open-vocabulary 3D instance localization, 3D point cloud understanding, click-based 3D object localization, hierarchical segmentation, and ablation studies. Our code is available at https://sy-ja.github.io/votesplat/

Figures

Figures reproduced from arXiv: 2506.22799 by the authors.

Figure 1
Figure 1. VoteSplat integrates 3DGS and Hough Voting for 3D scene understanding: (a) 3DGS retains its original novel view synthesis [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. Three main steps in VoteSplat pipeline: (a) We first deploy SAM to automatically generate segmentation masks for all instances [PITH_FULL_IMAGE:figures/full_fig_p004_2.png] view at source ↗
Figure 3
Figure 3. Since projection results in-depth information loss, pro [PITH_FULL_IMAGE:figures/full_fig_p005_3.png] view at source ↗
Figures from the paper (6 more)
Figure 4
Figure 4. Figure 4: Click-based 3D object selection and scene editing results. VoteSplat enables complete 3D object selection without issues of [PITH_FULL_IMAGE:figures/full_fig_p006_4.png]
Figure 5
Figure 5. Figure 5: Open-vocabulary 3D object selection on the LERF dataset. VoteSplat outperforms LangSplat and OpenGaussian in accurately [PITH_FULL_IMAGE:figures/full_fig_p007_5.png]
Figure 6
Figure 6. Figure 6: Comparison of point cloud feature visualizations. From left to right, the scenes correspond to [PITH_FULL_IMAGE:figures/full_fig_p008_6.png]
Figure 7
Figure 7. Figure 7: Using SAM, objects are divided into multiple parts, each assigned a 2D vote. After training, VoteSplat generates a corresponding [PITH_FULL_IMAGE:figures/full_fig_p008_7.png]
Figure 8
Figure 8. Figure 8: Instance segmentation in complex scenes. Compared with Open-Gaussian, VoteSplat can handle more complex scenarios, such [PITH_FULL_IMAGE:figures/full_fig_p008_8.png]
Figure 9
Figure 9. Figure 9: Comparison of ablation experiments. (a) shows 3D votes [PITH_FULL_IMAGE:figures/full_fig_p009_9.png]

Discussion (0). Sign in to comment.

Forward citations

Cited by 1 Pith paper

Reviewed papers in the Pith corpus that reference this work. Sorted by Pith novelty score. Full citation record

  1. ZeroSplat: Generalized Referring Segmentation in 3D Gaussian Splatting

    cs.CV 2026-07 conditional novelty 5.0 of 10

    ZeroSplat performs generalized referring segmentation in 3D Gaussian Splatting with no training or extra features, outperforming single-target baselines on two newly introduced benchmarks.

Reference graph

Works this paper leans on

43 extracted references · 31 canonical work pages · cited by 1 Pith paper

  1. [1]

    Dana H. Ballard. Generalizing the Hough transform to detect arbitrary shapes.Pattern recognition, 13(2):111–122, 1981. 3

    work page 1981

  2. [2]

    The 3D Hough transform for plane detection in point clouds: A review and a new accumulator design.3D Research, 2(2):1–13, 2011

    Dorit Borrmann, Jan Elseberg, Kai Lingemann, and Andreas N¨uchter. The 3D Hough transform for plane detection in point clouds: A review and a new accumulator design.3D Research, 2(2):1–13, 2011. 3

    work page 2011

  3. [3]

    Emerg- ing properties in self-supervised vision transformers

    Mathilde Caron, Hugo Touvron, Ishan Misra, Herv ´e J´egou, Julien Mairal, Piotr Bojanowski, and Armand Joulin. Emerg- ing properties in self-supervised vision transformers. InPro- ceedings of the IEEE/CVF international conference on com- puter vision, pages 9650–9660, 2021. 3

    2021

  4. [4]

    PGSR: Planar-based gaussian splat- ting for efficient and high-fidelity surface reconstruction

    Danpeng Chen, Hai Li, Weicai Ye, Yifan Wang, Weijian Xie, Shangjin Zhai, Nan Wang, Haomin Liu, Hujun Bao, and Guofeng Zhang. PGSR: Planar-based gaussian splat- ting for efficient and high-fidelity surface reconstruction. arXiv:2406.06521, 2024. 2

    arXiv 2024

  5. [5]

    Click-gaussian: Interactive segmenta- tion to any 3D gaussians

    Seokhun Choi, Hyeonseop Song, Jaechul Kim, Taehyeong Kim, and Hoseok Do. Click-gaussian: Interactive segmenta- tion to any 3D gaussians. InEuropean Conference on Com- puter Vision, pages 289–305. Springer, 2024. 2, 3, 4

    work page 2024

  6. [6]

    A Global Depth-Range-Free Multi-View Stereo Transformer Network with Pose Embedding

    Yitong Dong, Yijin Li, Zhaoyang Huang, Weikang Bian, Jingbo Liu, Hujun Bao, Zhaopeng Cui, Hongsheng Li, and Guofeng Zhang. A global depth-range-free multi-view stereo transformer network with pose embedding.arXiv preprint arXiv:2411.01893, 2024. 2

    work page Pith review arXiv 2024

  7. [7]

    Graspnet-1billion: A large-scale benchmark for general ob- ject grasping

    Hao-Shu Fang, Chenxi Wang, Minghao Gou, and Cewu Lu. Graspnet-1billion: A large-scale benchmark for general ob- ject grasping. InProceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 11444– 11453, 2020. 7

    work page 2020

  8. [8]

    Paul V .C. Hough. Machine analysis of bubble chamber pic- tures. InInternational Conference on High Energy Accel- erators and Instrumentation, CERN, 1959, pages 554–556,

    work page 1959

Show all 43 references
  1. [9]

    2D gaussian splatting for geometrically ac- curate radiance fields

    Binbin Huang, Zehao Yu, Anpei Chen, Andreas Geiger, and Shenghua Gao. 2D gaussian splatting for geometrically ac- curate radiance fields. InACM SIGGRAPH 2024 conference papers, pages 1–11, 2024. 2, 5

    2024

  2. [10]

    3D gaussian splatting for real-time radiance field rendering.ACM Trans

    Bernhard Kerbl, Georgios Kopanas, Thomas Leimk ¨uhler, and George Drettakis. 3D gaussian splatting for real-time radiance field rendering.ACM Trans. Graph., 42(4):139–1,

  3. [11]

    Lerf: Language embedded radiance fields

    Justin Kerr, Chung Min Kim, Ken Goldberg, Angjoo Kanazawa, and Matthew Tancik. Lerf: Language embedded radiance fields. InProceedings of the IEEE/CVF Interna- tional Conference on Computer Vision, pages 19729–19739,

  4. [12]

    Berg, Wan-Yen Lo, Piotr Doll ´ar, and Ross Girshick

    Alexander Kirillov, Eric Mintun, Nikhila Ravi, Hanzi Mao, Chloe Rolland, Laura Gustafson, Tete Xiao, Spencer White- head, Alexander C. Berg, Wan-Yen Lo, Piotr Doll ´ar, and Ross Girshick. Segment anything. InProceedings of the IEEE/CVF international conference on computer visi...

    2023

  5. [13]

    Orientation in- variant 3D object classification using Hough transform based methods

    Jan Knopp, Mukta Prasad, and Luc Van Gool. Orientation in- variant 3D object classification using Hough transform based methods. InProceedings of the ACM workshop on 3D object retrieval, pages 15–20, 2010. 3

    2010

  6. [14]

    Scene cut: Class-specific object detection and segmentation in 3D scenes

    Jan Knopp, Mukta Prasad, and Luc Van Gool. Scene cut: Class-specific object detection and segmentation in 3D scenes. In2011 International Conference on 3D Imag- ing, Modeling, Processing, Visualization and Transmission, pages 180–187. IEEE, 2011. 3

    2011

  7. [15]

    Robust object detection with interleaved categorization and segmen- tation.International journal of computer vision, 77:259– 289, 2008

    Bastian Leibe, Ale ˇs Leonardis, and Bernt Schiele. Robust object detection with interleaved categorization and segmen- tation.International journal of computer vision, 77:259– 289, 2008. 3

    2008

  8. [16]

    Instancegaussian: Appearance-semantic joint gaussian representation for 3D instance-level perception.arXiv:2411.19235, 2024

    Haijie Li, Yanmin Wu, Jiarui Meng, Qiankun Gao, Zhiyao Zhang, Ronggang Wang, and Jian Zhang. Instancegaussian: Appearance-semantic joint gaussian representation for 3D instance-level perception.arXiv:2411.19235, 2024. 3

    2024 arXiv

  9. [17]

    Weakly supervised 3D open- vocabulary segmentation.Advances in Neural Information Processing Systems, 36:53433–53456, 2023

    Kunhao Liu, Fangneng Zhan, Jiahui Zhang, Muyu Xu, Yingchen Yu, Abdulmotaleb El Saddik, Christian Theobalt, Eric Xing, and Shijian Lu. Weakly supervised 3D open- vocabulary segmentation.Advances in Neural Information Processing Systems, 36:53433–53456, 2023. 6

    2023

  10. [18]

    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 Pat- tern Recognition, pages 20654–20664, 2024. 2

    2024

  11. [19]

    hdbscan: Hierarchical density based clustering.J

    Leland McInnes, John Healy, and Steve Astels. hdbscan: Hierarchical density based clustering.J. Open Source Softw., 2(11):205, 2017. 5

    2017

  12. [20]

    Srinivasan, Matthew Tancik, Jonathan T

    Ben Mildenhall, Pratul P. Srinivasan, Matthew Tancik, Jonathan T. Barron, Ravi Ramamoorthi, and Ren Ng. Nerf: Representing scenes as neural radiance fields for view syn- thesis.Communications of the ACM, 65(1):99–106, 2021. 2, 3, 7

    2021

  13. [21]

    Qi, Hao Su, Kaichun Mo, and Leonidas J

    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,

  14. [22]

    Charles Ruizhongtai Qi, Li Yi, Hao Su, and Leonidas J. Guibas. Pointnet++: Deep hierarchical feature learning on point sets in a metric space.Advances in neural information processing systems, 30, 2017. 1

    2017

  15. [23]

    Qi, Or Litany, Kaiming He, and Leonidas J

    Charles R. Qi, Or Litany, Kaiming He, and Leonidas J. Guibas. Deep Hough voting for 3D object detection in point clouds. Inproceedings of the IEEE/CVF International Con- ference on Computer Vision, pages 9277–9286, 2019. 1, 3

    2019

  16. [24]

    Langsplat: 3D language gaussian splat- ting

    Minghan Qin, Wanhua Li, Jiawei Zhou, Haoqian Wang, and Hanspeter Pfister. Langsplat: 3D language gaussian splat- ting. InProceedings of the IEEE/CVF Conference on Com- puter Vision and Pattern Recognition, pages 20051–20060,

  17. [25]

    GLS: Geometry-aware 3D language gaussian splatting.arXiv preprint arXiv:2411.18066, 2024

    Jiaxiong Qiu, Liu Liu, Zhizhong Su, and Tianwei Lin. GLS: Geometry-aware 3D language gaussian splatting.arXiv preprint arXiv:2411.18066, 2024. 2

    2024 arXiv

  18. [26]

    You only look once: Unified, real-time object de- tection

    Joseph Redmon, Santosh Divvala, Ross Girshick, and Ali Farhadi. You only look once: Unified, real-time object de- tection. InProceedings of the IEEE conference on computer vision and pattern recognition, pages 779–788, 2016. 3

    2016

  19. [27]

    Faster R-CNN: Towards real-time object detection with re- gion proposal networks.IEEE transactions on pattern anal- ysis and machine intelligence, 39(6):1137–1149, 2016

    Shaoqing Ren, Kaiming He, Ross Girshick, and Jian Sun. Faster R-CNN: Towards real-time object detection with re- gion proposal networks.IEEE transactions on pattern anal- ysis and machine intelligence, 39(6):1137–1149, 2016. 3

    2016

  20. [28]

    Schonberger and Jan-Michael Frahm

    Johannes L. Schonberger and Jan-Michael Frahm. Structure- from-motion revisited. InProceedings of the IEEE con- ference on computer vision and pattern recognition, pages 4104–4113, 2016. 4

    2016

  21. [29]

    Language embedded 3D gaussians for open- vocabulary scene understanding

    Jin-Chuan Shi, Miao Wang, Hao-Bin Duan, and Shao- Hua Guan. Language embedded 3D gaussians for open- vocabulary scene understanding. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5333–5343, 2024. 2, 3, 4

    2024

  22. [30]

    Fast-Splat: Fast, ambiguity-free semantics transfer in gaussian splatting

    Ola Shorinwa, Jiankai Sun, and Mac Schwager. Fast-Splat: Fast, ambiguity-free semantics transfer in gaussian splatting. arXiv:2411.13753, 2024. 3

    2024 arXiv

  23. [31]

    Panoptic lifting for 3D scene understanding with neural fields

    Yawar Siddiqui, Lorenzo Porzi, Samuel Rota Bul ´o, Nor- man M ¨uller, Matthias Nießner, Angela Dai, and Peter Kontschieder. Panoptic lifting for 3D scene understanding with neural fields. InProceedings of the IEEE/CVF Con- ference on Computer Vision and Pattern Recognition, pa...

    2023

  24. [32]

    Depth-encoded Hough voting for joint object detection and shape recovery

    Min Sun, Gary Bradski, Bing-Xin Xu, and Silvio Savarese. Depth-encoded Hough voting for joint object detection and shape recovery. InEuropean Conference on Computer Vi- sion, pages 658–671. Springer, 2010. 3

    2010

  25. [33]

    Implicit shape models for object detection in 3D point clouds.ISPRS Annals of the Photogrammetry, Remote Sensing and Spatial Information Sciences, 1:179–184, 2012

    Alexander Velizhev, Roman Shapovalov, and Konrad Schindler. Implicit shape models for object detection in 3D point clouds.ISPRS Annals of the Photogrammetry, Remote Sensing and Spatial Information Sciences, 1:179–184, 2012. 3

    2012

  26. [34]

    Woodford, Minh-Tri Pham, Atsuto Maki, Frank Perbet, and Bj ¨orn Stenger

    Oliver J. Woodford, Minh-Tri Pham, Atsuto Maki, Frank Perbet, and Bj ¨orn Stenger. Demisting the Hough trans- form for 3D shape recognition and registration.International Journal of Computer Vision, 106:332–341, 2014. 3

    2014

  27. [35]

    Opengaussian: Towards point-level 3D gaussian-based open vocabulary understand- ing.arXiv:2406.02058, 2024

    Yanmin Wu, Jiarui Meng, Haijie Li, Chenming Wu, Yahao Shi, Xinhua Cheng, Chen Zhao, Haocheng Feng, Errui Ding, Jingdong Wang, and Jian Zhang. Opengaussian: Towards point-level 3D gaussian-based open vocabulary understand- ing.arXiv:2406.02058, 2024. 2, 3

    2024 arXiv

  28. [36]

    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–

  29. [37]

    Gaussian- dreamer: Fast generation from text to 3D gaussian splatting with point cloud priors.arXiv:2310.08529, 2023

    Taoran Yi, Jiemin Fang, Guanjun Wu, Lingxi Xie, Xiaopeng Zhang, Wenyu Liu, Qi Tian, and Xinggang Wang. Gaussian- dreamer: Fast generation from text to 3D gaussian splatting with point cloud priors.arXiv:2310.08529, 2023. 2

    2023 arXiv

  30. [38]

    Differentiable surface splatting for point-based geometry processing.ACM Transactions On Graphics (TOG), 38(6):1–14, 2019

    Wang Yifan, Felice Serena, Shihao Wu, Cengiz ¨Oztireli, and Olga Sorkine-Hornung. Differentiable surface splatting for point-based geometry processing.ACM Transactions On Graphics (TOG), 38(6):1–14, 2019. 3

    2019

  31. [39]

    Mip-splatting: Alias-free 3D gaussian splat- ting

    Zehao Yu, Anpei Chen, Binbin Huang, Torsten Sattler, and Andreas Geiger. Mip-splatting: Alias-free 3D gaussian splat- ting. InProceedings of the IEEE/CVF Conference on Com- puter Vision and Pattern Recognition, pages 19447–19456,

  32. [40]

    Gaussian opacity fields: Efficient and compact surface reconstruction in unbounded scenes.arXiv:2404.10772, 2024

    Zehao Yu, Torsten Sattler, and Andreas Geiger. Gaussian opacity fields: Efficient and compact surface reconstruction in unbounded scenes.arXiv:2404.10772, 2024. 2

    2024 arXiv

  33. [41]

    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 (CVPR), pages 14114–14124, 2025. 2

    2025

  34. [42]

    Shuaifeng Zhi, Tristan Laidlow, Stefan Leutenegger, and An- drew J. Davison. In-place scene labelling and understanding with implicit scene representation. InProceedings of the IEEE/CVF International Conference on Computer Vision, pages 15838–15847, 2021. 2

    2021

  35. [179]

    Springer, 2024. 2, 3, 6

    2024

Pith tools

Reviewed August 6, 2026 · model on record in the stance chip above.