Pith. sign in

REVIEW 3 major objections 5 minor 35 references

OpenSplat3D: Open-Vocabulary 3D Instance Segmentation using Gaussian Splatting

T0 review · 3 major / 5 minor · reviewed 2026-08-07 · deepseek-v4-flash

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

Pith's one-line read OpenSplat3D extends 3D Gaussian Splatting with per-Gaussian instance features and a variance regularization loss, achieving open-vocabulary 3D instance segmentation that outperforms OpenGaussian and Segment3D on LERF and ScanNet++…

desk verdict The ScanNet++ results are real, but the headline LERF-OVS win over OpenGaussian is confounded by a swapped VLM, so the abstract overclaims. read the letter →

arxiv 2506.07697 v1 pith:OR3CNJEQ submitted 2025-06-09 cs.CV

Jens Piekenbrinck , Christian Schmidt , Alexander Hermans , Narunas Vaskevicius , Timm Linder , Bastian Leibe This is my paper
classification cs.CV
keywords open-vocabulary3DinstancesegmentationGaussianSplattingSegmentAnythingModelcontrastivelearningvarianceregularizationlanguageembeddingsScanNet++zero-shot
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

OpenSplat3D claims that a 3D Gaussian Splatting scene can be turned into an open-vocabulary instance segmenter using only 2D supervision: SAM masks during training and a vision-language model at query time. The paper extends each Gaussian with a low-dimensional instance feature, trained by a contrastive loss and a new variance regularization that keeps features from smearing along rendering rays. On the LERF-mask, LERF-OVS, and ScanNet++ benchmarks it reports consistent gains over OpenGaussian and Segment3D, and it narrows the gap to fully supervised 3D segmentation. If true, this makes 3D instance segmentation practical for arbitrary scenes without manual 3D annotations.

What carries the argument

The central object is an instance feature field: each Gaussian carries a view-independent embedding $f_n \in \mathbb{R}^8$, rendered via the same $\alpha$-compositing as color. Supervision is a prototype contrastive loss that pulls pixels of a SAM mask toward their mean feature and pushes mask prototypes apart with margin $\gamma = 1$, plus a variance regularization $L_{\text{var}}$ that penalizes per-pixel variance of the rendered feature map along the ray, computed efficiently inside the CUDA rasterizer. At inference, HDBSCAN clusters the 8D embeddings into instances, and each instance's language embedding is the mean over top-5 views and 3 zoom levels of vision-language model crops.

What would settle it

Render a scene with two different SAM mask assignments for the same object across views (one mask in one view, two masks in another) and compare instance AP against a run with consistent masks: the majority-vote assumption predicts a measurable drop, and if no drop occurs, the contrastive loss is more robust than the paper assumes.

Watch

Extended reading notes

Core claim

The paper's central claim is that per-Gaussian instance features, supervised only by SAM masks and regularized by a ray-wise variance penalty, are enough to recover coherent 3D instances that can be named by language. Concretely, OpenSplat3D reports LERF-OVS mIoU 59.70 against OpenGaussian's 38.36, LERF-mask mean mIoU 84.0, class-agnostic ScanNet++ AP 24.5 against Segment3D's 20.2, and AP50 41.7 against 30.9, all with a single fixed hyperparameter set. The authors attribute the gains to the variance loss, which prevents alpha-composited feature blending from dissolving object boundaries, and they show that disabling it sharply reduces mIoU. The remaining gap to fully supervised SGIFormer is attributed mainly to SAM's tendency to over-segment.

Load-bearing premise

The method depends on SAM's per-view masks being consistent enough across viewpoints that conflicting segmentations cancel out during joint optimization; if SAM systematically splits or merges objects differently from view to view, the learned features will not cluster into the true instances.

Editorial extensions

If this is right

  • Since no 3D labels are needed, the approach can scale to arbitrary indoor or outdoor scenes from posed image sequences alone.
  • The margin-based contrastive loss with a fixed hyperparameter set removes the per-scene tuning that OpenGaussian requires.
  • The variance loss provides a general fix for feature-splatting methods where alpha blending mixes features from different objects during rendering.
  • The remaining gap to fully supervised SGIFormer is attributed mostly to SAM over-segmentation, so better mask proposals should directly improve instance AP.
  • Per-instance language embeddings enable open-vocabulary queries without optimizing per-Gaussian CLIP features, avoiding the need for scene-specific auto-encoder compression.

Reading between the lines

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

  • If SAM over-segmentation is indeed the bottleneck, a testable extension is to feed instance-aware prompts or self-refined masks back into the optimization loop, which should improve ScanNet++ AP without changing the loss formulation.
  • The method's reliance on majority voting across views implies that performance may degrade in scenes with heavy occlusion and few viewpoints; a stress-test benchmark with sparse views would reveal this limit.
  • The variance regularization could be applied to other per-Gaussian attributes, such as semantic features in language-splatting style fields, to reduce cross-object bleeding in those settings as well.
  • Averaging language embeddings across crops may underperform on very small or thin objects; weighting the aggregation by the rendered instance silhouette could improve retrieval accuracy.
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

3 major / 5 minor

Summary. The paper proposes OpenSplat3D, a 3D Gaussian Splatting-based method for open-vocabulary 3D instance segmentation. It extends each Gaussian with a low-dimensional instance feature, trained with a contrastive loss using SAM masks from individual views, plus a new variance regularization loss that penalizes feature variance along rendering rays. After optimization, HDBSCAN clusters the Gaussians into instances, and a per-instance language embedding is computed from top-K views and multiple zoom levels using a VLM (MasQCLIP). The method is evaluated on LERF-mask, LERF-OVS, and ScanNet++ validation, reporting state-of-the-art results against OpenGaussian, Segment3D, and others.

Significance. If substantiated, the results are a strong empirical contribution: the method integrates instance learning and open-vocabulary language alignment in a single 3DGS framework, the variance loss is a simple and apparently effective regularizer, and the evaluation spans three benchmarks. The paper also ships a plausible ablation for the variance loss on LERF-mask. However, the central comparative claims currently rest on single runs and on comparisons that do not hold the language encoder fixed, so the significance is conditional on additional evidence.

major comments (3)
  1. [Section 4.1, Table 2] The LERF-OVS comparison does not hold the language encoder fixed. OpenSplat3D uses MasQCLIP, chosen specifically for mask-conditioned object embeddings, while OpenGaussian uses a different CLIP-style VLM. Because LERF-OVS is a text-to-instance retrieval task, the choice of VLM can dominate the mIoU; the 59.70 vs 38.36 margin cannot be attributed to the proposed instance feature learning or variance regularization without an ablation that keeps the VLM identical across methods. The only variance-loss ablation (Fig. 5) is on LERF-mask, not LERF-OVS, so the statement 'We attribute this to our novel variance loss' in Section 4.1 is unsupported for the benchmark used in the headline claim.
  2. [All experimental tables (Tables 1-4)] All reported results are single runs without error bars, multiple seeds, or significance tests. The abstract and conclusion claim that the method 'significantly outperforms' OpenGaussian and Segment3D, but no statistical support is provided. Given the number of hyperparameters (loss weights, margin, top-K, HDBSCAN parameters) and the reported sensitivity of clustering, an assessment of variance across runs is needed to make the comparative claims credible.
  3. [Section 4.2, Tables 3-4 and abstract] The abstract claims significant outperformance over both OpenGaussian and Segment3D, but OpenGaussian is never evaluated on ScanNet++ (Tables 3-4). The only direct comparison to OpenGaussian is on LERF-OVS, which suffers from the VLM confound described above. To support the general claim, either a ScanNet++ comparison with OpenGaussian under the same protocol or a qualification of the claim to the LERF-OVS benchmark is needed.
minor comments (5)
  1. [Section 3.2.3, Eq. (12)] The variance formula in Eq. (12) applies the square f_n^2 and F(p)^2 elementwise, but this is not stated; please clarify that the operations are per-dimension and the l2 norm in Eq. (13) is over the d-dimensional variance vector.
  2. [Section 4, Implementation Details] The increased densification gradient threshold is not quantified; please state the value used.
  3. [Section 2] The claim that OpenSplat3D 'converges in fewer than half the iterations required by OpenGaussian' is made without evidence (e.g., a convergence plot); please either substantiate or remove.
  4. [Figure 5] The plot shows mIoU vs λvar but the marker at λvar=0 is not explicitly annotated as the no-variance baseline; adding a dashed line or label would improve readability.
  5. [Section 4.1, LERF-OVS protocol] The paragraph describing the OpenGaussian protocol says 'Selected Gaussians are rendered to binary masks from multiple viewpoints without occlusion handling'; it would help to state explicitly that this is the same protocol used for all methods in Table 2.

Circularity Check

0 steps flagged · score 1.0 of 10

No significant circularity: the 3D instance predictions are not built from the evaluation targets; the lone self-citation (Mask3D) is not load-bearing.

full rationale

OpenSplat3D is an empirical pipeline that combines external 2D foundation models (SAM masks, MasQCLIP language embeddings) with a 3DGS feature field optimized by the RGB loss, a margin-based contrastive loss, and the proposed variance loss (Eqs. 4-14). The final instance IDs are obtained by HDBSCAN clustering of the learned 8D per-Gaussian features rather than by reading off the SAM masks, and open-vocabulary labels are assigned by matching per-instance VLM embeddings to text embeddings, so the outputs are not equal to the training inputs by construction. The LERF-mask evaluation uses Grounding DINO and SAM only to select the queried instances; the reported mIoU is computed against the benchmark masks. No parameter is fitted to the evaluation subsets and then reported as a prediction; all loss weights and margins are fixed across datasets. The only self-citation is Mask3D [23], which appears in related work and as a closed-set comparison baseline in Table 4; it does not carry the argument, so it is not load-bearing. The paper's own caveats about SAM over-segmentation and cross-view inconsistency are robustness limitations, not circularity. The LERF-OVS comparison uses a different VLM for OpenSplat3D (MasQCLIP) than for OpenGaussian, which is a fairness/confound concern for the headline gain rather than a circularity. Overall, the derivation chain is self-contained against external benchmarks; score 1.

Assumptions & free parameters 8 free parameters · 5 assumptions · 0 invented entities

The central claim rests on standard 3DGS rendering machinery, SAM mask quality, HDBSCAN clustering behavior, and VLM embedding quality. No new physical entities are introduced. Several hyperparameters are chosen by hand but kept fixed across datasets, which is reasonable, but HDBSCAN and densification details are unspecified and therefore load-bearing unknowns.

free parameters (8)
  • Instance feature dimension d = 8
    Dimension of per-Gaussian instance embedding; chosen by hand and affects clustering and memory.
  • Contrastive negative margin gamma = 1.0
    Minimum squared distance between prototype pairs; directly shapes feature separation.
  • Loss weights lambda_inst2d, lambda_var, w_pos, w_neg = 0.1, 0.5, 1.0, 1.0
    Balance RGB, contrastive, and variance terms; fixed across all experiments.
  • Top-k selected views for language embedding = 5
    Number of most visible views used to compute per-instance language embedding.
  • Zoom levels L and expansion ratio = 3, 0.3
    Hierarchical crops for VLM embedding averaging.
  • HDBSCAN parameters = not specified
    Clustering parameters are not reported, so a major source of result variation is undocumented.
  • Increased densification gradient threshold = not specified
    The paper states the threshold is increased to prevent excessive Gaussian growth but does not give the value.
  • Number of frames per ScanNet++ scene = 300
    Uniformly sampled frames used per scene; affects supervision coverage and runtime.
assumptions (5)
  • domain assumption The 3DGS differentiable rendering pipeline correctly maps per-Gaussian features to 2D feature maps via alpha compositing (Eq. 5).
    Section 3.2 relies on this to supervise the 3D instance features with 2D masks.
  • domain assumption SAM masks provide sufficiently consistent instance-level supervision across viewpoints despite being generated independently per image.
    Section 3.2.1 states masks are not required to be consistent; the method assumes joint optimization acts as majority voting.
  • domain assumption HDBSCAN clustering on 8D instance embeddings yields meaningful and coherent 3D instances.
    Section 3.2.1 uses HDBSCAN to determine cluster structure and identify outliers without reporting parameters.
  • domain assumption Averaged MasQCLIP embeddings from top-k views and L zoom levels represent instance-level semantics well enough for open-vocabulary retrieval.
    Section 3.2.2 computes per-instance language embeddings by averaging VLM features from crops; this assumes the average captures the object's semantics.
  • ad hoc to paper Minimizing feature variance along rendering rays reduces blending artifacts and improves clusterability.
    Section 3.2.3 introduces the variance loss as a novel regularizer; its effectiveness is empirically shown but not derived from first principles.

how reviews work

0 comments
Cite this review

Pith. "Pith review of OpenSplat3D: Open-Vocabulary 3D Instance Segmentation using Gaussian Splatting." pith.science (2026). https://pith.science/paper/OR3CNJEQ

@misc{pith2026250607697,
  author       = {Pith},
  title        = {Pith review of: OpenSplat3D: Open-Vocabulary 3D Instance Segmentation using Gaussian Splatting},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/OR3CNJEQ}},
  note         = {Machine review of arXiv:2506.07697}
}
read the original abstract

3D Gaussian Splatting (3DGS) has emerged as a powerful representation for neural scene reconstruction, offering high-quality novel view synthesis while maintaining computational efficiency. In this paper, we extend the capabilities of 3DGS beyond pure scene representation by introducing an approach for open-vocabulary 3D instance segmentation without requiring manual labeling, termed OpenSplat3D. Our method leverages feature-splatting techniques to associate semantic information with individual Gaussians, enabling fine-grained scene understanding. We incorporate Segment Anything Model instance masks with a contrastive loss formulation as guidance for the instance features to achieve accurate instance-level segmentation. Furthermore, we utilize language embeddings of a vision-language model, allowing for flexible, text-driven instance identification. This combination enables our system to identify and segment arbitrary objects in 3D scenes based on natural language descriptions. We show results on LERF-mask and LERF-OVS as well as the full ScanNet++ validation set, demonstrating the effectiveness of our approach.

Figures

Figures reproduced from arXiv: 2506.07697 by the authors.

Figure 1
Figure 1. OpenSplat3D jointly optimizes 3D geometry and 3D [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. Overview of our proposed pipeline. On the left are the training inputs: posed RGB-images, a coarse SfM point cloud for [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗
Figure 3
Figure 3. Language query results on the LERF-mask dataset. The left side shows example RGB training images for reference. The right [PITH_FULL_IMAGE:figures/full_fig_p007_3.png] view at source ↗
Figures from the paper (2 more)
Figure 4
Figure 4. Figure 4: Qualitative Results on four scenes of the ScanNet++ validation split. Top row showing the mesh of the scan in a top-down [PITH_FULL_IMAGE:figures/full_fig_p008_4.png]
Figure 5
Figure 5. Figure 5: Influence of the variance loss, evaluated on the LERF [PITH_FULL_IMAGE:figures/full_fig_p008_5.png]

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

35 extracted references · 33 canonical work pages

  1. [1]

    Ricardo J. G. B. Campello, Davoud Moulavi, and Joerg Sander. Density-based clustering based on hierarchical den- sity estimates. In Advances in Knowledge Discovery and Data Mining, 2013. 4

    work page 2013

  2. [2]

    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. In ICCV, 2021. 1, 2

    work page 2021

  3. [3]

    Click-Gaussian: Interactive segmen- tation to any 3D gaussians

    Seokhun Choi, Hyeonseop Song, Jaechul Kim, Taehyeong Kim, and Hoseok Do. Click-Gaussian: Interactive segmen- tation to any 3D gaussians. In ECCV, 2024. 2, 4

    work page 2024

  4. [4]

    Chang, Manolis Savva, Maciej Hal- ber, Thomas Funkhouser, and Matthias Nießner

    Angela Dai, Angel X. Chang, Manolis Savva, Maciej Hal- ber, Thomas Funkhouser, and Matthias Nießner. ScanNet: Richly-annotated 3D reconstructions of indoor scenes. In CVPR, 2017. 2

    work page 2017

  5. [5]

    OpenNeRF: Open set 3D neural scene segmentation with pixel-wise features and rendered novel views

    Francis Engelmann, Fabian Manhardt, Michael Niemeyer, Keisuke Tateno, Marc Pollefeys, and Federico Tombari. OpenNeRF: Open set 3D neural scene segmentation with pixel-wise features and rendered novel views. InICLR, 2024. 2

    work page 2024

  6. [6]

    Efficient graph-based image segmentation

    Pedro F Felzenszwalb and Daniel P Huttenlocher. Efficient graph-based image segmentation. IJCV, 59, 2004. 7, 8

    work page 2004

  7. [7]

    EgoLifter: Open-world 3D seg- mentation for egocentric perception

    Qiao Gu, Zhaoyang Lv, Duncan Frost, Simon Green, Julian Straub, and Chris Sweeney. EgoLifter: Open-world 3D seg- mentation for egocentric perception. In ECCV, 2024. 2, 4

    work page 2024

  8. [8]

    OccuSeg: Occupancy-aware 3D instance segmentation

    Lei Han, Tian Zheng, Lan Xu, and Lu Fang. OccuSeg: Occupancy-aware 3D instance segmentation. In CVPR,

    work page

Show all 35 references
  1. [9]

    Segment3D: Learning fine-grained class-agnostic 3D segmentation without manual labels

    Rui Huang, Songyou Peng, Ayca Takmaz, Federico Tombari, Marc Pollefeys, Shiji Song, Gao Huang, and Francis Engel- mann. Segment3D: Learning fine-grained class-agnostic 3D segmentation without manual labels. In ECCV, 2025. 1, 2, 5, 7, 8

    2025

  2. [10]

    3D Gaussian Splatting for Real-Time Radiance Field Rendering

    Bernhard Kerbl, Georgios Kopanas, Thomas Leimk ¨uhler, and George Drettakis. 3D Gaussian Splatting for Real-Time Radiance Field Rendering. ACM TOG, 2023. 1, 2, 3, 5

    2023

  3. [11]

    LERF: Language embed- ded radiance fields

    Justin Kerr, Chung Min Kim, Ken Goldberg, Angjoo Kanazawa, and Matthew Tancik. LERF: Language embed- ded radiance fields. In CVPR, 2023. 2, 4, 6

    2023

  4. [12]

    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. In CVPR, 2023. 1, 2, 4, 5

    2023

  5. [13]

    Top-down beats bottom-up in 3D instance segmentation

    Maksim Kolodiazhnyi, Anna V orontsova, Anton Konushin, and Danila Rukhovich. Top-down beats bottom-up in 3D instance segmentation. In WACV, 2024. 2

    2024

  6. [14]

    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. In CVPR, 2024. 2

    2024

  7. [15]

    Grounding DINO: Marrying DINO with grounded pre-training for open-set object detection

    Shilong Liu, Zhaoyang Zeng, Tianhe Ren, Feng Li, Hao Zhang, Jie Yang, Qing Jiang, Chunyuan Li, Jianwei Yang, Hang Su, et al. Grounding DINO: Marrying DINO with grounded pre-training for open-set object detection. In ECCV, 2024. 6

    2024

  8. [16]

    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. In ECCV, 2020. 2

    2020

  9. [17]

    Open3DIS: Open-vocabulary 3D instance segmentation with 2D mask guidance

    Phuc Nguyen, Tuan Duc Ngo, Evangelos Kalogerakis, Chuang Gan, Anh Tran, Cuong Pham, and Khoi Nguyen. Open3DIS: Open-vocabulary 3D instance segmentation with 2D mask guidance. In CVPR, 2024. 2, 8

    2024

  10. [18]

    DINOv2: Learning robust visual features without supervi- sion

    Maxime Oquab, Timoth ´ee Darcet, Th ´eo Moutakanni, Huy V o, Marc Szafraniec, Vasil Khalidov, Pierre Fernandez, Daniel Haziza, Francisco Massa, Alaaeldin El-Nouby, et al. DINOv2: Learning robust visual features without supervi- sion. TMLR, 2024. 1, 2

    2024

  11. [19]

    OpenScene: 3D scene understanding with open vocabular- ies

    Songyou Peng, Kyle Genova, Chiyu Jiang, Andrea Tagliasacchi, Marc Pollefeys, Thomas Funkhouser, et al. OpenScene: 3D scene understanding with open vocabular- ies. In CVPR, 2023. 2

    2023

  12. [20]

    LangSplat: 3D language gaussian splat- ting

    Minghan Qin, Wanhua Li, Jiawei Zhou, Haoqian Wang, and Hanspeter Pfister. LangSplat: 3D language gaussian splat- ting. In CVPR, 2024. 2, 4, 6

    2024

  13. [21]

    Learn- ing transferable visual models from natural language super- vision

    Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learn- ing transferable visual models from natural language super- vision. In ICML, 2021. 1, 2

    2021

  14. [22]

    Ma- chine learning in Python: Main developments and technol- ogy trends in data science, machine learning, and artificial intelligence

    Sebastian Raschka, Joshua Patterson, and Corey Nolet. Ma- chine learning in Python: Main developments and technol- ogy trends in data science, machine learning, and artificial intelligence. Information, 11(4):193, 2020. 5

    2020

  15. [23]

    Mask3D: Mask trans- former for 3D semantic instance segmentation

    Jonas Schult, Francis Engelmann, Alexander Hermans, Or Litany, Siyu Tang, and Bastian Leibe. Mask3D: Mask trans- former for 3D semantic instance segmentation. In ICRA,

  16. [24]

    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. In CVPR, 2024. 6

    2024

  17. [25]

    Silva, Mahtab Dahaghin, Matteo Toso, and Alessio Del Bue

    Myrna C. Silva, Mahtab Dahaghin, Matteo Toso, and Alessio Del Bue. Contrastive gaussian clustering for weakly super- vised 3D scene segmentation. In ICPR, 2024. 2, 4, 6

    2024

  18. [26]

    Superpoint transformer for 3D scene instance segmentation

    Jiahao Sun, Chunmei Qing, Junpeng Tan, and Xiangmin Xu. Superpoint transformer for 3D scene instance segmentation. AAAI, 2023. 2

    2023

  19. [27]

    Sumner, Marc Pollefeys, Federico Tombari, and Francis Engelmann

    Ayc ¸a Takmaz, Elisabetta Fedele, Robert W. Sumner, Marc Pollefeys, Federico Tombari, and Francis Engelmann. Open- Mask3D: Open-V ocabulary 3D Instance Segmentation. In NeurIPS, 2023. 1, 2, 5, 7, 8

    2023

  20. [28]

    Bovik, H.R

    Zhou Wang, A.C. Bovik, H.R. Sheikh, and E.P. Simoncelli. Image quality assessment: from error visibility to structural similarity. TIP, 13(4):600–612, 2004. 4

    2004

  21. [29]

    OpenGaussian: Towards point-level 3D gaussian-based open vocabulary understand- ing

    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. In NeurIPS, 2024. 2, 5, 6, 7

    2024

  22. [30]

    MasQCLIP for open-vocabulary universal image segmenta- tion

    Xin Xu, Tianyi Xiong, Zheng Ding, and Zhuowen Tu. MasQCLIP for open-vocabulary universal image segmenta- tion. In ICCV, 2023. 2, 5

    2023

  23. [31]

    SAM3D: Segment anything in 3D scenes

    Yunhan Yang, Xiaoyang Wu, Tong He, Hengshuang Zhao, and Xihui Liu. SAM3D: Segment anything in 3D scenes. In ICCV Workshops, 2023. 1, 8

    2023

  24. [32]

    SGI- Former: Semantic-guided and geometric-enhanced inter- leaving transformer for 3D instance segmentation

    Lei Yao, Yi Wang, Moyun Liu, and Lap-Pui Chau. SGI- Former: Semantic-guided and geometric-enhanced inter- leaving transformer for 3D instance segmentation. TCSVT,

  25. [33]

    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. In ECCV, 2024. 2, 4, 6

    2024

  26. [34]

    ScanNet++: A high-fidelity dataset of 3D indoor scenes

    Chandan Yeshwanth, Yueh-Cheng Liu, Matthias Nießner, and Angela Dai. ScanNet++: A high-fidelity dataset of 3D indoor scenes. In CVPR, 2023. 1, 2, 6

    2023

  27. [35]

    Sigmoid loss for language image pre-training

    Xiaohua Zhai, Basil Mustafa, Alexander Kolesnikov, and Lucas Beyer. Sigmoid loss for language image pre-training. In ICCV, 2023. 1, 2

    2023

Pith tools

Reviewed August 7, 2026 · model on record in the stance chip above.