REVIEW 4 major objections 5 minor 1 cited by
CCL-LGS: Contrastive Codebook Learning for 3D Language Gaussian Splatting
T0 review · 4 major / 5 minor · reviewed 2026-08-07 · deepseek-v4-flash
Pith's one-line read The paper claims that view-consistent mask tracking plus contrastive codebook learning produces state-of-the-art 3D open-vocabulary segmentation in Gaussian splatting.
desk verdict Useful incremental advance on 3D language fields with a real LERF win, but the contrastive labels are instance tracks rather than semantic classes, and the SOTA claim needs correcting. read the letter →
The pith
A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.
The reading
What carries the argument
The central object is the Contrastive Codebook Learning (CCL) module: a learnable codebook $\mathcal{T}=\{T_j\}_{j=1}^N$ of prototype vectors $T_j \in \mathbb{R}^d$ that structures the feature space before 3D supervision. It does three jobs at once: a matching loss $L_{\max}=1-\cos(F_i,T_{j^*})$ binds each CLIP feature to its nearest prototype; a pull loss $L_{\text{pull}}=1-\cos(T_{j_i},T_{j_k})$ collapses features whose tracked masks share a category; and a push loss $L_{\text{push}}=\operatorname{ReLU}(\cos(T_{j_i},T_{j_k})-m)$ separates features whose categories differ. These losses are applied only to masks with confident category assignments ($y_i \neq -1$), so the codebook functions as a denoiser that absorbs the noise of imperfect, partially occluded masks. The cleaned codebook indices, not the raw CLIP vectors, are what supervise the 3D Gaussian semantic field.
What would settle it
Run CCL-LGS on one LERF scene with the association threshold varied over $\{0.3, 0.4, 0.5, 0.6, 0.7\}$ and with SAM2 tracking replaced by manually corrected oracle masks; if mIoU swings sharply with the threshold, or the oracle version substantially beats the tracker version, then the claimed robustness to imperfect masks comes from the tracker and threshold rather than from the contrastive codebook.
Extended reading notes
Core claim
On its own terms, the paper's discovery is that supervision for 3D Gaussian semantic fields should be organized as discrete, view-consistent categories rather than raw CLIP embeddings. After extracting SAM masks at two aggregated scales (subpart+part and whole+part), the method uses SAM2 to propagate first-frame masks to all views and assigns each view mask the category of the propagated mask whose IoU exceeds 0.5. A learnable codebook of $N$ prototype vectors acts as a bottleneck: each mask feature is bound to its nearest prototype through a matching loss, same-category prototypes are pulled together, and different-category prototypes are pushed apart with margin $m=0.7$. The resulting index maps supervise the low-dimensional features stored on 3D Gaussians through a cross-entropy loss. The claim is that this pipeline yields more compact, discriminative, and view-consistent semantic features, and that this is what produces the reported gains over prior methods.
Load-bearing premise
The load-bearing premise is that SAM2's propagated masks and the IoU > 0.5 rule assign the correct semantic category to each view mask; if tracking drifts or two objects fall into one mask, the pull and push losses reinforce wrong clusters and corrupt the 3D semantic field.
Editorial extensions
If this is right
- If the central claim holds, 3D language Gaussian fields become more stable under occlusion and blur; the largest per-scene gain on LERF is Waldo Kitchen, where CCL-LGS reaches 67.1 mIoU versus 54.8 for 3D VL-GS.
- Because supervision is a codebook index rather than a dense CLIP vector, inference can retrieve one prototype per pixel and answer text queries by cosine similarity, enabling language-based 3D object selection and editing without re-extracting CLIP features.
- The ablation attributes the gain to both contrastive terms working together: pull loss raises the baseline from 56.4 to 59.5, push loss to 60.4, and both together to 65.6 on LERF.
- On datasets with mild occlusion and little viewpoint variation, the advantage narrows and reverses on 3D-OVS (95.2 versus 96.9), so the method's value is conditional on how inconsistent the 2D supervision actually is.
Reading between the lines
- The codebook-denoising pattern could be transferred to other 2D-to-3D feature lifting pipelines, including NeRF-based feature fields, wherever partial masks or patch features cause cross-view label noise; the paper does not test this.
- A direct stress test would be to swap SAM2 tracking identities for two visually similar objects in one scene and measure mIoU; a large drop would show that the contrastive losses cannot repair wrong category assignments, leaving the system's ceiling set by the tracker.
- The IoU 0.5 threshold is a categorical commitment that the paper does not analyze; sweeping it on LERF scenes would reveal where under-association (too many unmatched masks) or over-association (merged objects) begins.
- Unmatched masks (category $-1$) receive no contrastive supervision; treating them as an active-learning signal to re-query with different prompts or viewpoints is a natural extension that the paper leaves implicit.
Editorial analysis
A structured set of objections, weighed in public.
Referee Report
Summary. The paper proposes CCL-LGS, a framework for 3D language Gaussian splatting that aims to produce view-consistent semantic supervision. The pipeline first generates two-scale SAM masks, then propagates the first-frame masks with SAM2 and associates masks across views by IoU matching, and finally trains a codebook with contrastive pull/push losses to refine CLIP features. The refined features are used to supervise a 3D Gaussian semantic field through cross-entropy on codebook indices. Experiments are reported on LERF and 3D-OVS datasets using mIoU. On LERF, the method reports 65.6 average mIoU versus 62.0 for 3D VL-GS; on 3D-OVS, it reports 95.2 versus 96.9 for 3D VL-GS. Ablations on LERF show that both pull and push losses improve over the baseline, and a supplementary section reports efficiency comparisons and a scale-aggregation ablation. The conclusion explicitly acknowledges that imperfections in SAM and SAM2 remain a limitation.
Significance. The cross-view semantic inconsistency problem addressed by the paper is genuine and relevant for 3D open-vocabulary segmentation. The contrastive codebook module is a simple and potentially reusable idea, and the LERF result of 65.6 mIoU is a meaningful improvement over the 62.0 reported for 3D VL-GS if it is reproducible. The paper includes a per-scene ablation showing that both loss terms contribute, an efficiency analysis in the supplement, and an explicit limitation statement about SAM/SAM2 imperfection. However, the central claim that the method enforces category-level discriminability is not directly supported by the construction, because the contrastive labels are SAM2 instance identities rather than semantic categories. The reported SOTA claim is also weakened by the 3D-OVS result, by the absence of error bars, and by unspecified codebook-size and threshold choices. If the category-level claim can be repaired or re-framed as instance-level consistency with supporting analysis, the contribution would be solid. The paper ships no code, but that is not by itself a blocker.
major comments (4)
- [Sec. 3.3, Eqs. (6)-(7)] The contrastive labels y_i in {1,...,K,-1} are instance identities, not semantic categories: K is the number of SAM masks in the first frame, and the IoU matching step only propagates these instance identities across frames. Consequently, L_pull clusters features of the same tracked instance, while L_push separates features of different tracked instances even when those instances belong to the same semantic class. The paper's claims in the abstract and Sec. 1 that CCL 'preserves category discriminability' and enforces 'inter-class distinctiveness' are therefore not established by the construction. This is load-bearing because Eq. (9) supervises the 3D field using these codebook labels, and Eq. (10) normalizes over all codebook prototypes; if same-class instances are pushed apart, multi-instance queries may be penalized. The authors should either change the contrastive labeling to semantic categories (e.g., by grouping instances with CLIP or text labels) or provide a direct analysis showing that instance-level contrastive training does not harm category-level discrimination, for example by reporting per-query mIoU separated into single-instance and multi-instance queries.
- [Abstract and Sec. 4.2, Table 3] The abstract claims that CCL-LGS 'outperforms previous state-of-the-art methods' without qualification, but Table 3 shows that on 3D-OVS the method achieves 95.2 average mIoU versus 96.9 for 3D VL-GS, i.e., it underperforms the same baseline on that benchmark. The Sec. 4.2 text acknowledges this and attributes it to the small/simple nature of 3D-OVS, but the unqualified abstract and concluding 'state-of-the-art' statements are contradicted by the paper's own table. The SOTA claim should be restricted to LERF or to scenes with occlusion, blur, and viewpoint variation, and the 3D-OVS deficit should be discussed with concrete evidence rather than only a post-hoc explanation.
- [Implementation Details and Sec. 3.3] The codebook size N is a central design parameter of the method, but its value is never reported: Sec. 3.3 defines T = {T_j}_{j=1}^N and Eq. (9) uses N as the number of output classes, yet the Implementation Details paragraph gives d_f = 8 but not N. Similarly, the IoU threshold of 0.5 in Sec. 3.3, the contrastive weights lambda_pull = lambda_push = 0.25, and the margin m = 0.7 are fixed without any sensitivity study. Without the codebook size and without any analysis of how N or the IoU threshold affect mIoU, the method is not fully reproducible and the robustness of the reported 65.6 result cannot be assessed. The authors should state N explicitly and add a sensitivity experiment for N and for the IoU association threshold, even if only on one LERF scene.
- [Sec. 3.3 and Sec. 5 (Limitations)] The method's load-bearing premise is that SAM2 tracking and the IoU > 0.5 rule correctly assign the same semantic identity to masks across views. The paper does not provide any tracking failure analysis, nor does it quantify how often masks are assigned label -1 or how tracker errors propagate through the pull and push losses into the 3D semantic field. The limitation paragraph in Sec. 5 states only that imperfect SAM and SAM2 masks still affect results; it does not examine the specific failure mode of wrong cross-view associations. This matters because, as the paper itself notes, a wrong association will cause the contrastive losses to cluster or separate features according to incorrect labels, and Eq. (9) then bakes those errors into the 3D Gaussians. The authors should add a quantitative study of mask-association accuracy (e.g., fraction of frames where IoU > 0.5, examples of tracking failures) and show how such failures affect final mIoU.
minor comments (5)
- [Sec. 3.2, Eq. (3)] The notation CLIP(I_t ⊙ M_i(v)) is unclear: M_i(v) appears to denote a mask region, but the expression should specify whether the entire masked image is passed to CLIP and then the feature is assigned to every pixel in the mask, or whether a per-pixel crop is used. Please define the operation precisely.
- [Fig. 1 caption] The caption of Fig. 1 says 'Quantitative comparison', but the figure shows qualitative visualizations of segmentation maps, not quantitative plots. Please correct the caption to 'Qualitative comparison'.
- [Sec. 3.4, Eq. (10)] In Eq. (10), the denominator sums over s ∈ T, but T is defined in Sec. 3.3 as the set of prototypes, while elsewhere the prototypes are indexed as T_j. Please align the notation so that the summation index and the set are consistently defined.
- [Sec. 3.2] The phrase 'a uniform 32×32 point prompt' is ambiguous; it should say 'a uniform grid of 32×32 points' if that is the intended meaning, because a point prompt in SAM is usually a single point.
- [Sec. 4.2, Table 3] The text says the method achieves 'comparable performance' on 3D-OVS, but the gap to 3D VL-GS is 1.7 points on average and 3.2 points on the Sofa scene. Please quantify this statement, for example by reporting the per-scene gaps explicitly or by providing a statistical comparison.
Circularity Check
No material circularity: the reported mIoU is measured against external ground-truth masks, and the per-scene training does not fit the benchmark answers.
full rationale
The derivation chain is self-contained with respect to the evaluated benchmarks. SAM/SAM2 provide masks and propagated instance labels, CLIP provides the feature and text embeddings, the CCL module trains a codebook with pull and push losses on those self-generated labels, and the 3D Gaussians are supervised with cross-entropy against the resulting index maps (Eqs. 4-9). None of these stages fits the LERF or 3D-OVS ground-truth masks used for the reported mIoU in Tables 1 and 3; the ground truth is consulted only at evaluation time. The paper does not rely on a load-bearing self-citation: all compared baselines are external prior works, and no uniqueness theorem from the authors is invoked to forbid alternative designs. The only arguable concern is not circularity but construct validity: the labels y_i are propagated SAM2 instance identities rather than semantic categories, so the push loss in Eq. (7) may separate same-class instances, and the paper's own limitation statement concedes that imperfect SAM/SAM2 masks still affect results. That weakness does not make the benchmark numbers equivalent to the method's inputs by construction, so the appropriate finding is no significant circularity.
Assumptions & free parameters
free parameters (5)
- contrastive loss weights lambda_pull, lambda_push =
0.25 each
- push loss margin m =
0.7
- IoU threshold for mask-to-category assignment =
0.5
- codebook size N =
not reported
- semantic feature dimension df =
8
assumptions (5)
- domain assumption CLIP embeddings computed on masked image regions are reliable semantic supervision for 3D fields.
- domain assumption SAM2 zero-shot tracking correctly propagates object masks across views, and IoU > 0.5 between a propagated mask and a frame mask implies same semantic category.
- domain assumption Two mask scales (subpart+part and whole+part) are sufficient to resolve semantic ambiguity.
- ad hoc to paper A fixed-size codebook of N prototypes can represent all categories in a scene.
- domain assumption Different objects of the same category should share one codebook entry, while distinct categories should be separated, despite viewpoint-dependent appearance.
invented entities (1)
-
Learned codebook prototypes T_j (j=1..N)
Cite this review
Pith. "Pith review of CCL-LGS: Contrastive Codebook Learning for 3D Language Gaussian Splatting." pith.science (2026). https://pith.science/paper/5ECNGFT2
@misc{pith2026250520469,
author = {Pith},
title = {Pith review of: CCL-LGS: Contrastive Codebook Learning for 3D Language Gaussian Splatting},
year = {2026},
howpublished = {\url{https://pith.science/paper/5ECNGFT2}},
note = {Machine review of arXiv:2505.20469}
}
read the original abstract
Recent advances in 3D reconstruction techniques and vision-language models have fueled significant progress in 3D semantic understanding, a capability critical to robotics, autonomous driving, and virtual/augmented reality. However, methods that rely on 2D priors are prone to a critical challenge: cross-view semantic inconsistencies induced by occlusion, image blur, and view-dependent variations. These inconsistencies, when propagated via projection supervision, deteriorate the quality of 3D Gaussian semantic fields and introduce artifacts in the rendered outputs. To mitigate this limitation, we propose CCL-LGS, a novel framework that enforces view-consistent semantic supervision by integrating multi-view semantic cues. Specifically, our approach first employs a zero-shot tracker to align a set of SAM-generated 2D masks and reliably identify their corresponding categories. Next, we utilize CLIP to extract robust semantic encodings across views. Finally, our Contrastive Codebook Learning (CCL) module distills discriminative semantic features by enforcing intra-class compactness and inter-class distinctiveness. In contrast to previous methods that directly apply CLIP to imperfect masks, our framework explicitly resolves semantic conflicts while preserving category discriminability. Extensive experiments demonstrate that CCL-LGS outperforms previous state-of-the-art methods. Our project page is available at https://epsilontl.github.io/CCL-LGS/.
Figures
Figures from the paper (8 more)
Forward citations
Cited by 1 Pith paper
-
Disentangling concept semantics via multilingual averaging in Sparse Autoencoders
The abstract claims multilingual averaging of Gemma Scope activations aligns with ontology ground truth better than any single language, but the provided full text is an unrelated paper and contains no supporting evidence.
Reference graph
Works this paper leans on
-
[1]
Text-to-3d using gaussian splatting
Zilong Chen, Feng Wang, Yikai Wang, and Huaping Liu. Text-to-3d using gaussian splatting. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 21401–21412, 2024. 2
work page 2024
-
[2]
Xmem: Long- term video object segmentation with an atkinson-shiffrin memory model
Ho Kei Cheng and Alexander G Schwing. Xmem: Long- term video object segmentation with an atkinson-shiffrin memory model. In European Conference on Computer Vi- sion, pages 640–658. Springer, 2022. 3
work page 2022
-
[3]
Differentiable surface render- ing via non-differentiable sampling
Forrester Cole, Kyle Genova, Avneesh Sud, Daniel Vla- sic, and Zhoutong Zhang. Differentiable surface render- ing via non-differentiable sampling. In Proceedings of the IEEE/CVF International Conference on Computer Vision , pages 6088–6097, 2021. 2
work page 2021
-
[4]
Hybrid neural rendering for large-scale scenes with motion blur
Peng Dai, Yinda Zhang, Xin Yu, Xiaoyang Lyu, and Xiao- juan Qi. Hybrid neural rendering for large-scale scenes with motion blur. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , pages 154–164,
-
[5]
Panoptic nerf: 3d-to-2d label transfer for panoptic urban scene segmentation
Xiao Fu, Shangzhan Zhang, Tianrun Chen, Yichong Lu, Lanyun Zhu, Xiaowei Zhou, Andreas Geiger, and Yiyi Liao. Panoptic nerf: 3d-to-2d label transfer for panoptic urban scene segmentation. In 2022 International Conference on 3D Vision (3DV), pages 1–11. IEEE, 2022. 3
2022
-
[6]
Gaussianflow: Splatting gaussian dynamics for 4d content creation
Quankai Gao, Qiangeng Xu, Zhe Cao, Ben Mildenhall, Wen- chao Ma, Le Chen, Danhang Tang, and Ulrich Neumann. Gaussianflow: Splatting gaussian dynamics for 4d content creation. arXiv preprint arXiv:2403.12365, 2024. 2
arXiv 2024
-
[7]
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 Trans. Graph., 42(4):139–1,
-
[8]
Lerf: Language embedded radiance fields
Justin Kerr, Chung Min Kim, Ken Goldberg, Angjoo Kanazawa, and Matthew Tancik. Lerf: Language embedded radiance fields. In Proceedings of the IEEE/CVF Interna- tional Conference on Computer Vision, pages 19729–19739,
Show all 35 references
-
[9]
Autosplat: Constrained gaussian splatting for autonomous driving scene reconstruction
Mustafa Khan, Hamidreza Fazlali, Dhruv Sharma, Tongtong Cao, Dongfeng Bai, Yuan Ren, and Bingbing Liu. Autosplat: Constrained gaussian splatting for autonomous driving scene reconstruction. arXiv preprint arXiv:2407.02598, 2024. 2
2024 arXiv
-
[10]
Kingma and Jimmy Ba
DiederikP. Kingma and Jimmy Ba. Adam: A method for stochastic optimization. arXiv: Learning,arXiv: Learning ,
-
[11]
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 Proceedings of the IEEE/CVF international con- ference on computer vision , pages 4015–4026, 20...
2023
-
[12]
Decomposing nerf for editing via feature field distil- lation
Sosuke Kobayashi, Eiichi Matsumoto, and Vincent Sitz- mann. Decomposing nerf for editing via feature field distil- lation. Advances in neural information processing systems , 35:23311–23330, 2022. 3
2022
-
[13]
Gart: Gaussian articulated template mod- els
Jiahui Lei, Yufu Wang, Georgios Pavlakos, Lingjie Liu, and Kostas Daniilidis. Gart: Gaussian articulated template mod- els. In Proceedings of the IEEE/CVF conference on com- puter vision and pattern recognition , pages 19876–19887,
-
[14]
Language-driven semantic seg- mentation
Boyi Li, Kilian Q Weinberger, Serge Belongie, Vladlen Koltun, and Ren ´e Ranftl. Language-driven semantic seg- mentation. arXiv preprint arXiv:2201.03546, 2022. 2
2022 arXiv
-
[15]
Weakly supervised 3d open- vocabulary segmentation, 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, 2023. 6
2023
-
[16]
Nerf: Representing scenes as neural radiance fields for view syn- thesis
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
2021
-
[17]
Dinov2: Learning robust visual features without supervision
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 supervision. arXiv preprint arXiv:2304.07193, 2023. 3
2023 arXiv
-
[18]
3d vision-language gaussian splatting
Qucheng Peng, Benjamin Planche, Zhongpai Gao, Meng Zheng, Anwesa Choudhuri, Terrence Chen, Chen Chen, and Ziyan Wu. 3d vision-language gaussian splatting. arXiv preprint arXiv:2410.07577, 2024. 2, 3, 5, 6
2024 arXiv
-
[19]
Dynamic point fields
Sergey Prokudin, Qianli Ma, Maxime Raafat, Julien Valentin, and Siyu Tang. Dynamic point fields. In Proceed- ings of the IEEE/CVF International Conference on Com- puter Vision (ICCV), pages 7964–7976, 2023. 2
2023
-
[20]
Langsplat: 3d language gaussian splatting
Minghan Qin, Wanhua Li, Jiawei Zhou, Haoqian Wang, and Hanspeter Pfister. Langsplat: 3d language gaussian splatting. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 20051–20060, 2024. 2, 3, 4, 5, 6
2024
-
[21]
Ad- vancing extended reality with 3d gaussian splatting: Innova- tions and prospects
Shi Qiu, Binzhu Xie, Qixuan Liu, and Pheng-Ann Heng. Ad- vancing extended reality with 3d gaussian splatting: Innova- tions and prospects. In 2025 IEEE International Conference on Artificial Intelligence and eXtended and Virtual Reality (AIxVR), pages 203–208. IEEE, 2025. 2
2025
-
[22]
Goi: Find 3d gaussians of interest with an optimizable open-vocabulary semantic-space hyperplane
Yansong Qu, Shaohui Dai, Xinyang Li, Jianghang Lin, Liu- juan Cao, Shengchuan Zhang, and Rongrong Ji. Goi: Find 3d gaussians of interest with an optimizable open-vocabulary semantic-space hyperplane. In Proceedings of the 32nd ACM International Conference on Multimedia, pages ...
-
[23]
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. In International conference on machine learning, ...
2021
-
[24]
Sam 2: Segment anything in images and videos
Nikhila Ravi, Valentin Gabeur, Yuan-Ting Hu, Ronghang Hu, Chaitanya Ryali, Tengyu Ma, Haitham Khedr, Roman R¨adle, Chloe Rolland, Laura Gustafson, et al. Sam 2: Segment anything in images and videos. arXiv preprint arXiv:2408.00714, 2024. 3, 5
2024 arXiv
-
[25]
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 Proceedings of the 9 IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5333–5343, 2024. 2, 3, 5, 6
2024
-
[26]
Dreamgaussian: Generative gaussian splatting for effi- cient 3d content creation
Jiaxiang Tang, Jiawei Ren, Hang Zhou, Ziwei Liu, and Gang Zeng. Dreamgaussian: Generative gaussian splatting for effi- cient 3d content creation. arXiv preprint arXiv:2309.16653,
-
[27]
Neural feature fusion fields: 3d distillation of self-supervised 2d image representations
Vadim Tschernezki, Iro Laina, Diane Larlus, and Andrea Vedaldi. Neural feature fusion fields: 3d distillation of self-supervised 2d image representations. In 2022 Inter- national Conference on 3D Vision (3DV) , pages 443–453. IEEE, 2022. 3
2022
-
[28]
Rip-nerf: Learning rotation-invariant point-based neural radiance field for fine-grained editing and compositing
Yuze Wang, Junyi Wang, Yansong Qu, and Yue Qi. Rip-nerf: Learning rotation-invariant point-based neural radiance field for fine-grained editing and compositing. In Proceedings of the 2023 ACM international conference on multimedia re- trieval, pages 125–134, 2023. 2
2023
-
[29]
Rl-gsbridge: 3d gaussian splatting based real2sim2real method for robotic manipula- tion learning
Yuxuan Wu, Lei Pan, Wenhua Wu, Guangming Wang, Yanzi Miao, Fan Xu, and Hesheng Wang. Rl-gsbridge: 3d gaussian splatting based real2sim2real method for robotic manipula- tion learning. arXiv preprint arXiv:2409.20291, 2024. 2
2024 arXiv
-
[30]
Point- nerf: Point-based neural radiance fields
Qiangeng Xu, Zexiang Xu, Julien Philip, Sai Bi, Zhixin Shu, Kalyan Sunkavalli, and Ulrich Neumann. Point- nerf: Point-based neural radiance fields. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 5438–5448, 2022. 2
2022
-
[31]
Gs-slam: Dense visual slam with 3d gaussian splatting
Chi Yan, Delin Qu, Dan Xu, Bin Zhao, Zhigang Wang, Dong Wang, and Xuelong Li. Gs-slam: Dense visual slam with 3d gaussian splatting. In Proceedings of the IEEE/CVF Con- ference on Computer Vision and Pattern Recognition, pages 19595–19604, 2024. 2
2024
-
[32]
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 European Conference on Computer Vision , pages 162–
-
[33]
Dino: Detr with improved denoising anchor boxes for end-to-end object detection
Hao Zhang, Feng Li, Shilong Liu, Lei Zhang, Hang Su, Jun Zhu, Lionel M Ni, and Heung-Yeung Shum. Dino: Detr with improved denoising anchor boxes for end-to-end object detection. arXiv preprint arXiv:2203.03605, 2022. 3
2022 arXiv
-
[34]
In-place scene labelling and understanding with implicit scene representation
Shuaifeng Zhi, Tristan Laidlow, Stefan Leutenegger, and An- drew J Davison. In-place scene labelling and understanding with implicit scene representation. In Proceedings of the IEEE/CVF International Conference on Computer Vision , pages 15838–15847, 2021. 3
2021
-
[35]
wood wall
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 Compu...
2024
Reviewed August 7, 2026 · model on record in the stance chip above.
Discussion (0). Sign in to comment.