Pith. sign in

REVIEW 3 major objections 4 minor 64 references

InstaScene: Towards Complete 3D Instance Decomposition and Reconstruction from Cluttered Scenes

T0 review · 3 major / 4 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 One pipeline extracts complete objects from cluttered 3D scenes

desk verdict A genuinely capable decomposition system with a plausible but under-evidenced completion claim; the paper deserves a serious referee, though the strong 'seamless alignment' language needs softening. read the letter →

arxiv 2507.08416 v2 pith:PZ6RWR3G submitted 2025-07-11 cs.CV

Zesong Yang , Bangbang Yang , Wenqi Dong , Chenxuan Cao , Liyuan Cui , Yuewen Ma , Zhaopeng Cui , Hujun Bao This is my paper
classification cs.CV
keywords 3DinstancesegmentationGaussianSplattingscenedecompositionamodalcompletiongenerativecontrastivelearningin-situgenerationinstance-awarereconstruction
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

InstaScene aims to give machines a human-like ability to look at a cluttered 3D scene, pick out any object, and obtain that object as a complete, standalone 3D model—even the parts that were never visible. The paper argues that with the right supervision, a Gaussian Splatting scene can be decomposed reliably into arbitrary instances, and that a 3D generative model can complete each instance's hidden geometry and appearance while staying aligned with the original scene's scale, shape, and texture. If this holds, it closes the gap between whole-scene reconstruction and object-level generation, enabling realistic scene editing, robot manipulation, and simulation directly from captured scans.

What carries the argument

The central object is the spatial tracker: a set of Gaussian points that contributed to a mask's rasterization with transmittance above 0.5. These trackers are clustered across views using a view consensus rate, forming 3D instance masks; the masks then supervise a 16-dimensional instance feature field through contrastive losses in three regimes (single view, adjacent views, global 3D points). The completion side is driven by an omni-conditioned diffusion loop in which view conditions alternate per denoising step and known latent features are warped onto unseen views using rendered depths and surface normals, so the generative prior is anchored to observed geometry.

What would settle it

Render a completed instance from a viewpoint that was occluded in the input, and compare it against a held-out photo of the actual object seen from that angle; if the generated geometry or texture disagrees for objects outside the generative model's training distribution, the physical-alignment claim fails.

Watch

Extended reading notes

Core claim

InstaScene's central claim is that instance decomposition and complete reconstruction should be solved together, and that each stage can feed the other. The paper shows that tracing which Gaussians contributed to each 2D segmentation mask, then clustering these traces across views, yields reliable 3D instance masks that, combined with filtered 2D masks, train a highly distinctive feature field via spatial contrastive learning. For completion, it uses an in-situ generation pipeline that feeds all known observations and geometry cues as conditions to a pretrained 3D diffusion model, then jointly fine-tunes the instance's Gaussian model with source and generated views, producing objects that are complete and that match the physical scene.

Load-bearing premise

Completion quality rests on the pretrained 3D generative model having a prior that matches the true unseen parts of the object being completed.

Editorial extensions

If this is right

  • Objects are extracted as complete 3D models that can be placed back into the original scene with alignment to scale, geometry, and appearance.
  • The feature field learned with mutual 2D/3D supervision gives fine-grained instance segmentation even in cluttered scenes with repeated objects.
  • Completion and refinement happen jointly for geometry and texture, avoiding separate reconstruction stages.
  • The pipeline is open-set, requiring no CAD models or category-specific training for decomposition.

Reading between the lines

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

  • Beyond the paper, the tracing mechanism only needs a differentiable rasterizer, so the same decomposition logic could be reimplemented for other 3D representations if a suitable generative prior is available.
  • For objects far outside the generative prior's training distribution, 'complete' reconstructions may be plausible but wrong; robotics users would want a confidence signal before trusting the unseen parts.
  • The learned instance feature field could be paired with language embeddings to support text-based queries over the decomposed objects, an easy extension the authors do not explore.
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 / 4 minor

Summary. The manuscript proposes InstaScene, a pipeline for open-set instance decomposition and complete reconstruction in Gaussian Splatting scenes. It first clusters noisy 2D segmentation masks into 3D instance masks by tracing rasterized Gaussian contributions across views, then trains a per-Gaussian feature field with spatial contrastive losses that combine 2D intra-view, 2D cross-view, and global 3D mask supervision. For each decomposed incomplete object, an in-situ generation stage uses MVDFusion as a 3D generative prior, conditioned through alternated rendered views and geometry-aware feature warping, and jointly fine-tunes the object's 2D Gaussians with source and generated views. Experiments report large improvements over LangSplat and GSGrouping on LERF-Mask mIoU and over InstantMesh and SpaRP on Replica-CAD completion metrics, plus qualitative results on ZipNeRF.

Significance. The proposed combination of Gaussian-tracing mask clustering with contrastive feature learning is a sensible way to stabilise noisy 2D segmentation, and the reported segmentation gains are substantial if reproducible. The in-situ generation stage is a principled way to reuse all available observations and geometry to guide a generic image-to-3D prior; if the completion results generalize beyond the Replica-CAD objects, the system would be practically useful for scene editing and robot manipulation. The paper does not ship code or trained models, and several evaluation choices leave the strength of the evidence uncertain. However, the main idea is clear, the ablations support the value of the individual components, and the writing is generally accessible.

major comments (3)
  1. [Sec. 3.3, Eq. (8), and Table 3] The claim that completed instances 'seamlessly align with the real world' is not supported for unseen parts. The denoising step in Eq. (8) averages noise predictions conditioned only on rendered views of the incomplete instance, and the geometry-aware warping projects features from source views onto visible pixels of target views; pixels invisible in every source view receive no geometry constraint and are therefore synthesized purely from the MVDFusion prior. The quantitative completion evaluation in Table 3 uses only Replica-CAD, a synthetic dataset, while the real-world ZipNeRF results are qualitative. There is no experiment with a held-out shape distribution or with real ground-truth geometry for cluttered objects, so the central 'arbitrary instances' claim remains unverified exactly in the regime where the generative prior could manufacture a plausible but incorrect object. The Conclusion should either add such an evaluation or explicitly list out-of-distribution static shapes as a limitation.
  2. [Sec. 4.1 and Table 1] The authors state that they 'manually re-annotate' the three LERF-Mask scenes with instance-level ground-truth masks. This creates a potential evaluation bias: the annotators are the method authors, and no annotation protocol, number of annotators, or inter-annotator agreement is reported. Since the mIoU numbers in Table 1 are the main quantitative evidence for the segmentation contribution, please replace or supplement these annotations with independent annotations, or at least report agreement statistics.
  3. [Tables 1-3] All quantitative tables report single runs without error bars, number of seeds, or significance tests. Given that the method has several thresholds (transmittance 0.5, visibility 30%, containment 80%, consensus 0.9, and tau_seg = 0.9 in Sec. 3.2), the reader cannot tell whether the large gaps over baselines are robust to these choices. Please provide variance over at least three runs, or a sensitivity analysis over the main thresholds.
minor comments (4)
  1. [Abstract and Sec. 1] The phrase 'arbitrary instances' is broader than the actual assumptions: the method relies on EntitySeg detecting the object and on static, opaque, non-reflective objects. Please qualify the scope in the abstract and introduction.
  2. [Sec. 3.2, Eq. (5)] The definitions of Nvis and Ncontain in Eq. (5) are given only in prose. Formal set-theoretic definitions of 'visible' and 'contained' for trackers would improve precision and reproducibility.
  3. [Table 3] Please clarify how 'Unknown' views are defined for the 2DGS 'Origin Recon' row, and why an incomplete model has a nonzero unknown-view PSNR. Also state explicitly whether the same generated target views are used for all compared methods.
  4. [Sec. 3.3] Several implementation details, such as the number of diffusion steps, the fine-tuning schedule, and the exact noise schedule for the latent warping, are deferred to the supplementary material. For a journal version, the main text should include at least a concise summary of these settings.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: decomposition and completion are validated against independent benchmarks, and the completion stage's reliance on an external generative prior is a correctness/scope risk, not a circular step.

full rationale

InstaScene's derivation chain has two main components, neither of which reduces to its own inputs. Scene decomposition (Sec. 3.2) clusters off-the-shelf EntitySeg masks by tracing Gaussian rasterization and view-consensus rates (Eq. 5), then trains a feature field with spatial contrastive losses (Eqs. 6-7). The output segmentation is not forced to equal the input 2D masks by construction: the clustered 3D masks filter and supervise a learned feature field, and results are evaluated against manually re-annotated GT on LERF-Mask and qualitatively on ZipNeRF. The completion stage (Sec. 3.3) conditions the external MVDFusion diffusion model on rendered views and geometry-aware warped features (Eq. 8); the unobserved regions are generated from that pretrained prior rather than derived from the scene. This is a heavy reliance on an external model, and for objects outside MVDFusion's training distribution the 'complete reconstruction' can be hallucinated rather than physically faithful. That is a genuine limitation and correctness risk, but it is not circularity: no fitted parameter is renamed as a prediction, and no equation defines the claimed output in terms of itself. Quantitative completion metrics (CD, F1, Volume IoU, PSNR/SSIM/LPIPS on Replica-CAD) compare against synthetic GT shapes; the 'unknown' views are viewpoints excluded from the conditioning set, not views reused as supervisory signal for the reported numbers. The paper does cite prior work by overlapping authors ([6], [56]), but only in related-work context, and these citations are not load-bearing for the central claim. The conclusion's limitation passage ('we cannot decompose dynamic, transparent or highly reflective objects') is a scope restriction, not a circularity admission. Overall, the derivation is self-contained against external benchmarks; score 0.

Assumptions & free parameters 8 free parameters · 6 assumptions · 0 invented entities

The paper introduces no new physical or mathematical entities. Its 'spatial trackers' and 'feature fields' are standard constructs in the 3D scene understanding literature. The method relies on a large set of hand-set thresholds and on pretrained models whose assumptions are not independently validated in the paper.

free parameters (8)
  • transmittance threshold for spatial tracker = 0.5
    Gaussians contributing to a mask are defined by transmittance exceeding 0.5; no sensitivity analysis is provided. Sec 3.2.
  • visibility threshold = 30%
    A tracker is visible in another frame if 30% of its points contribute to rasterization; no sensitivity analysis. Sec 3.2.
  • containment threshold = 80%
    A tracker is contained in another if 80% of its points appear; no sensitivity analysis. Sec 3.2.
  • view consensus rate threshold = 0.9
    Masks with consensus rate above 0.9 are merged as the same instance; no sensitivity analysis. Eq 5, Sec 3.2.
  • instance segmentation threshold = 0.9
    Cosine similarity threshold tau_seg=0.9 used to assign Gaussians to instances during inference; no sensitivity analysis. Sec 3.2.
  • number of target viewpoints = 16
    16 viewpoints centered around the object are selected for completion; value is asserted without justification. Sec 3.3.
  • feature dimension D = 16
    Each Gaussian is augmented with a 16-dimensional instance feature; no study of dimensionality. Sec 3.1.
  • contrastive loss weights lambda1, lambda2, lambda3 = not reported
    Equation 7 weights for intra-view, cross-view, and 3D contrastive losses; values are not given in the main text, implying they were tuned experimentally.
assumptions (6)
  • domain assumption 2D Gaussian Splatting provides an accurate enough reconstruction and surface geometry for downstream segmentation and feature warping.
    The entire pipeline operates on a 2DGS reconstruction; if the reconstruction is poor in visible regions, the geometry-aware feature warping in Sec 3.3 will propagate errors.
  • domain assumption EntitySeg provides reliable class-agnostic 2D instance masks across all scene objects.
    Mask clustering in Sec 3.2 starts from EntitySeg masks; any object never segmented in any view is invisible to the pipeline.
  • domain assumption The pretrained 3D diffusion model (MVDFusion) has a prior broad enough to generate the unseen regions of arbitrary real-world objects.
    Sec 3.3 relies on this prior for completion; if the object is out-of-distribution, completion is hallucination rather than reconstruction.
  • domain assumption View consensus rate is a reliable indicator of cross-view instance identity.
    Eq 5 and the 0.9 threshold in Sec 3.2 assume that high consensus implies the same object; this can fail for symmetric or repetitive objects that look similar from many views.
  • domain assumption Contrastive feature learning with cosine similarity produces separable instance features in cluttered scenes.
    The final segmentation in Sec 3.2 thresholds cosine similarity at 0.9; the method assumes the trained feature field is sufficiently discriminative.
  • standard math Standard math operations: alpha-blended rasterization (Eq 3-4), DBSCAN clustering, and the InfoNCE-style contrastive loss (Eq 6) are valid in this setting.
    These are standard tools; no new mathematical claims are made about them.

how reviews work

0 comments
Cite this review

Pith. "Pith review of InstaScene: Towards Complete 3D Instance Decomposition and Reconstruction from Cluttered Scenes." pith.science (2026). https://pith.science/paper/PZ6RWR3G

@misc{pith2026250708416,
  author       = {Pith},
  title        = {Pith review of: InstaScene: Towards Complete 3D Instance Decomposition and Reconstruction from Cluttered Scenes},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/PZ6RWR3G}},
  note         = {Machine review of arXiv:2507.08416}
}
read the original abstract

Humans can naturally identify and mentally complete occluded objects in cluttered environments. However, imparting similar cognitive ability to robotics remains challenging even with advanced reconstruction techniques, which models scenes as undifferentiated wholes and fails to recognize complete object from partial observations. In this paper, we propose InstaScene, a new paradigm towards holistic 3D perception of complex scenes with a primary goal: decomposing arbitrary instances while ensuring complete reconstruction. To achieve precise decomposition, we develop a novel spatial contrastive learning by tracing rasterization of each instance across views, significantly enhancing semantic supervision in cluttered scenes. To overcome incompleteness from limited observations, we introduce in-situ generation that harnesses valuable observations and geometric cues, effectively guiding 3D generative models to reconstruct complete instances that seamlessly align with the real world. Experiments on scene decomposition and object completion across complex real-world and synthetic scenes demonstrate that our method achieves superior decomposition accuracy while producing geometrically faithful and visually intact objects.

Figures

Figures reproduced from arXiv: 2507.08416 by the authors.

Figure 1
Figure 1. InstaScene allows users to pick up and decompose arbitrary instances from cluttered environments, while automatically recon￾structing them into complete objects with intact geometry and appearance that align with the physical world. Abstract Humans can naturally identify and mentally complete oc￾cluded objects in cluttered environments. However, impart￾ing similar cognitive ability to robotics remains challenging ev… view at source ↗
Figure 2
Figure 2. Motivation of spatial contrastive learning with mu￾tual guidance. In complex scenes, naively supervising feature field with noisy 2D segmentation masks results in indistinct fea￾tures (see (a)), and only using 3D masks from spatial trackers would result in sparse Gaussian points (see (b)). Observed that the former provides dense features while the latter offers a robust reference, we leverage the interplay to mutual… view at source ↗
Figure 3
Figure 3. System Overview. Given a reconstructed Gaussian Splatting scene, our method first clusters and filters 2D segmentation masks by tracing the rasterization of Gaussian Splatting, which yields 2D and 3D instance masks. Then, we use spatial contrastive learning with mask supervision to train a feature field that achieves high-quality scene decomposition. Finally, for each decomposed incomplete object, we conduct an in-s… view at source ↗
Figures from the paper (6 more)
Figure 4
Figure 4. Figure 4: In-Situ Generation vs. Generic Image-to-3D. We per￾form different instance-level complete reconstruction approaches on the baby carriage and put it back on the scene. Generic image￾to-3D methods [52, 53] struggle to maintain consistent reconstruc￾tion with the original…
Figure 5
Figure 5. Figure 5: Comparison on LERF-Mask Dataset. We compare the segmentation results and visualized features (with PCA) of GSGroup￾ing [58], LangSplat [36] and ours. Our method learns highly distinguishable features and achieves the most precise instance segmentation. Methods Figurine…
Figure 6
Figure 6. Figure 6: Comparison on ZipNeRF Dataset. The baselines yield highly noisy segmentation results in such complex scenes, while our method achieves fine-grained instance-level segmentation. We also show the segmented Gaussian points of our method. the viewing-ray-mesh intersection …
Figure 7
Figure 7. Figure 7: Comparisons on Replica-CAD Dataset [45]. Instances extracted from the decomposition suffer from incomplete reconstruction, our method preserves the quality of the originally visible regions while achieving most plausible recovery of unknown regions. Methods PSNR↑ SSIM↑…
Figure 8
Figure 8. Figure 8: Comparison with different generation methods. We show the complete reconstruction results for each method on diverse instances. Our method not only achieves faithful completion but also maintains consistency with the original scene rendering. Single-View Input Single-V…
Figure 9
Figure 9. Figure 9: Ablation studies of the omni-conditioned completion from in-situ generation. Na¨ıvely alternating the condition views resolves the unrealistic predictions of single-view input, and the geometry-aware feature warping further enhances the consistency of the generated vie…

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

64 extracted references · 34 canonical work pages

  1. [1]

    Scan2cad: Learning cad model alignment in rgb-d scans

    Armen Avetisyan, Manuel Dahnert, Angela Dai, Manolis Savva, Angel X Chang, and Matthias Nießner. Scan2cad: Learning cad model alignment in rgb-d scans. In Proceed- ings of the IEEE/CVF Conference on computer vision and pattern recognition, pages 2614–2623, 2019. 2

    work page 2019

  2. [2]

    Zip-nerf: Anti-aliased grid-based neural radiance fields

    Jonathan T Barron, Ben Mildenhall, Dor Verbin, Pratul P Srinivasan, and Peter Hedman. Zip-nerf: Anti-aliased grid-based neural radiance fields. In Proceedings of the IEEE/CVF International Conference on Computer Vision , pages 19697–19705, 2023. 6, 7

    work page 2023

  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. In Pro- ceedings of the IEEE/CVF international conference on com- puter vision, pages 9650–9660, 2021. 3

    2021

  4. [4]

    Shapenet: An information-rich 3d model repository

    Angel X Chang, Thomas Funkhouser, Leonidas Guibas, Pat Hanrahan, Qixing Huang, Zimo Li, Silvio Savarese, Manolis Savva, Shuran Song, Hao Su, et al. Shapenet: An information-rich 3d model repository. arXiv preprint arXiv:1512.03012, 2015. 2

    arXiv 2015

  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. In European Conference on Com- puter Vision, pages 289–305. Springer, 2025. 3, 4, 6

    work page 2025

  6. [6]

    Coin3d: Controllable and interactive 3d assets generation with proxy- guided conditioning

    Wenqi Dong, Bangbang Yang, Lin Ma, Xiao Liu, Liyuan Cui, Hujun Bao, Yuewen Ma, and Zhaopeng Cui. Coin3d: Controllable and interactive 3d assets generation with proxy- guided conditioning. In ACM SIGGRAPH 2024 Conference Papers, pages 1–10, 2024. 3

    work page 2024

  7. [7]

    A density-based algorithm for discovering clusters in large spatial databases with noise

    Martin Ester, Hans-Peter Kriegel, J ¨org Sander, Xiaowei Xu, et al. A density-based algorithm for discovering clusters in large spatial databases with noise. In kdd, pages 226–231,

  8. [8]

    3d-front: 3d furnished rooms with layouts and semantics

    Huan Fu, Bowen Cai, Lin Gao, Ling-Xiao Zhang, Jiaming Wang, Cao Li, Qixun Zeng, Chengyue Sun, Rongfei Jia, Bin- qiang Zhao, et al. 3d-front: 3d furnished rooms with layouts and semantics. In Proceedings of the IEEE/CVF Interna- tional Conference on Computer Vision, pages 10933–10942,

    work page

Show all 64 references
  1. [9]

    Lrm: Large reconstruction model for single image to 3d

    Yicong Hong, Kai Zhang, Jiuxiang Gu, Sai Bi, Yang Zhou, Difan Liu, Feng Liu, Kalyan Sunkavalli, Trung Bui, and Hao Tan. Lrm: Large reconstruction model for single image to 3d. arXiv preprint arXiv:2311.04400, 2023. 1, 2

    2023 arXiv

  2. [10]

    Mvd-fusion: Single-view 3d via depth-consistent multi-view generation

    Hanzhe Hu, Zhizhuo Zhou, Varun Jampani, and Shubham Tulsiani. Mvd-fusion: Single-view 3d via depth-consistent multi-view generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , pages 9698–9707, 2024. 3, 5, 7, 8

    2024

  3. [11]

    Oˆ 2-recon: Com- pleting 3d reconstruction of occluded objects in the scene with a pre-trained 2d diffusion model

    Yubin Hu, Sheng Ye, Wang Zhao, Matthieu Lin, Yuze He, Yu-Hui Wen, Ying He, and Yong-Jin Liu. Oˆ 2-recon: Com- pleting 3d reconstruction of occluded objects in the scene with a pre-trained 2d diffusion model. In Proceedings of the AAAI Conference on Artificial Intelligence, pag...

    2024

  4. [12]

    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. In ACM SIGGRAPH 2024 Conference Papers, pages 1–11, 2024. 3, 7

    2024

  5. [13]

    Graspsplats: Efficient manipulation with 3d feature splatting

    Mazeyu Ji, Ri-Zhao Qiu, Xueyan Zou, and Xiaolong Wang. Graspsplats: Efficient manipulation with 3d feature splatting. arXiv preprint arXiv:2409.02084, 2024. 1

    2024 arXiv

  6. [14]

    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,

  7. [15]

    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,

  8. [16]

    Garfield: Group anything with radiance fields

    Chung Min Kim, Mingxuan Wu, Justin Kerr, Ken Gold- berg, Matthew Tancik, and Angjoo Kanazawa. Garfield: Group anything with radiance fields. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 21530–21539, 2024. 3, 4

    2024

  9. [17]

    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, 2023. 3

    2023

  10. [18]

    A review of robot learning for manipulation: Challenges, rep- resentations, and algorithms

    Oliver Kroemer, Scott Niekum, and George Konidaris. A review of robot learning for manipulation: Challenges, rep- resentations, and algorithms. Journal of machine learning research, 22(30):1–82, 2021. 1

    2021

  11. [19]

    Rico: Regularizing the unobservable for indoor compositional reconstruction

    Zizhang Li, Xiaoyang Lyu, Yuanyuan Ding, Mengmeng Wang, Yiyi Liao, and Yong Liu. Rico: Regularizing the unobservable for indoor compositional reconstruction. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 17761–17771, 2023. 2

    2023

  12. [20]

    Part123: part-aware 3d reconstruction from a single-view image

    Anran Liu, Cheng Lin, Yuan Liu, Xiaoxiao Long, Zhiyang Dou, Hao-Xiang Guo, Ping Luo, and Wenping Wang. Part123: part-aware 3d reconstruction from a single-view image. In ACM SIGGRAPH 2024 Conference Papers, pages 1–12, 2024. 3, 4

    2024

  13. [21]

    Lasa: Instance reconstruction from real scans using a large-scale aligned shape annotation dataset

    Haolin Liu, Chongjie Ye, Yinyu Nie, Yingfan He, and Xi- aoguang Han. Lasa: Instance reconstruction from real scans using a large-scale aligned shape annotation dataset. In Pro- ceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 20454–20464, 2...

    2024

  14. [22]

    Weakly supervised 3d open- vocabulary segmentation

    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

  15. [23]

    Zero-1-to- 3: Zero-shot one image to 3d object

    Ruoshi Liu, Rundi Wu, Basile Van Hoorick, Pavel Tok- makov, Sergey Zakharov, and Carl V ondrick. Zero-1-to- 3: Zero-shot one image to 3d object. In Proceedings of the IEEE/CVF international conference on computer vision, pages 9298–9309, 2023. 3

    2023

  16. [24]

    Syncdreamer: Gen- erating multiview-consistent images from a single-view im- age

    Yuan Liu, Cheng Lin, Zijiao Zeng, Xiaoxiao Long, Lingjie Liu, Taku Komura, and Wenping Wang. Syncdreamer: Gen- erating multiview-consistent images from a single-view im- age. arXiv preprint arXiv:2309.03453, 2023. 3, 6, 7

    2023 arXiv

  17. [25]

    Infusion: Inpainting 3d gaussians via learn- ing depth completion from diffusion prior

    Zhiheng Liu, Hao Ouyang, Qiuyu Wang, Ka Leong Cheng, Jie Xiao, Kai Zhu, Nan Xue, Yu Liu, Yujun Shen, and Yang Cao. Infusion: Inpainting 3d gaussians via learn- ing depth completion from diffusion prior. arXiv preprint arXiv:2404.11613, 2024. 3

    2024 arXiv

  18. [26]

    Ovir-3d: Open-vocabulary 3d in- stance retrieval without training on 3d data

    Shiyang Lu, Haonan Chang, Eric Pu Jing, Abdeslam Boular- ias, and Kostas Bekris. Ovir-3d: Open-vocabulary 3d in- stance retrieval without training on 3d data. In Conference on Robot Learning, pages 1610–1620. PMLR, 2023. 4

    2023

  19. [27]

    Repaint: Inpainting using denoising diffusion probabilistic models

    Andreas Lugmayr, Martin Danelljan, Andres Romero, Fisher Yu, Radu Timofte, and Luc Van Gool. Repaint: Inpainting using denoising diffusion probabilistic models. In Proceed- ings of the IEEE/CVF conference on computer vision and pattern recognition, pages 11461–11471, 2022. 3

    2022

  20. [28]

    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. 1, 2

    2021

  21. [29]

    Spin-nerf: Multiview segmentation and perceptual inpainting with neural radiance fields

    Ashkan Mirzaei, Tristan Aumentado-Armstrong, Konstanti- nos G Derpanis, Jonathan Kelly, Marcus A Brubaker, Igor Gilitschenski, and Alex Levinshtein. Spin-nerf: Multiview segmentation and perceptual inpainting with neural radiance fields. In Proceedings of the IEEE/CVF Conferen...

  22. [30]

    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 Proceedings of the IEEE/CVF Con- ference on Computer Vision and Pattern Recognition, pages 4018–402...

    2024

  23. [31]

    Decompositional neu- ral scene reconstruction with generative diffusion prior

    Junfeng Ni, Yu Liu, Ruijie Lu, Zirui Zhou, Song-Chun Zhu, Yixin Chen, and Siyuan Huang. Decompositional neu- ral scene reconstruction with generative diffusion prior. In CVPR, 2025. 3

    2025

  24. [32]

    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

  25. [33]

    Openscene: 3d scene understanding with open vocabularies

    Songyou Peng, Kyle Genova, Chiyu Jiang, Andrea Tagliasacchi, Marc Pollefeys, Thomas Funkhouser, et al. Openscene: 3d scene understanding with open vocabularies. In Proceedings of the IEEE/CVF conference on computer vi- sion and pattern recognition, pages 815–824, 2023. 1

    2023

  26. [34]

    Dreamfusion: Text-to-3d using 2d diffusion

    Ben Poole, Ajay Jain, Jonathan T Barron, and Ben Milden- hall. Dreamfusion: Text-to-3d using 2d diffusion. arXiv preprint arXiv:2209.14988, 2022. 1, 2

    2022 arXiv

  27. [35]

    High-quality entity segmentation

    Lu Qi, Jason Kuen, Weidong Guo, Tiancheng Shen, Jiuxiang Gu, Jiaya Jia, Zhe Lin, and Ming-Hsuan Yang. High-quality entity segmentation. arXiv preprint arXiv:2211.05776, 2022. 4, 6

    2022 arXiv

  28. [36]

    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. 3, 5, 6

    2024

  29. [37]

    Feature splatting: Language-driven physics-based scene syn- thesis and editing

    Ri-Zhao Qiu, Ge Yang, Weijia Zeng, and Xiaolong Wang. Feature splatting: Language-driven physics-based scene syn- thesis and editing. arXiv preprint arXiv:2404.01223, 2024. 3

    2024 arXiv

  30. [38]

    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

  31. [39]

    Structure- from-motion revisited

    Johannes L Schonberger and Jan-Michael Frahm. Structure- from-motion revisited. In Proceedings of the IEEE con- ference on computer vision and pattern recognition , pages 4104–4113, 2016. 1

    2016

  32. [40]

    Pixelwise view selection for un- structured multi-view stereo

    Johannes Lutz Sch ¨onberger, Enliang Zheng, Marc Pollefeys, and Jan-Michael Frahm. Pixelwise view selection for un- structured multi-view stereo. In European Conference on Computer Vision (ECCV), 2016. 2

    2016

  33. [41]

    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 IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5333–5343, 2024. 3

    2024

  34. [42]

    Zero123++: a single image to consistent multi-view dif- fusion base model

    Ruoxi Shi, Hansheng Chen, Zhuoyang Zhang, Minghua Liu, Chao Xu, Xinyue Wei, Linghao Chen, Chong Zeng, and Hao Su. Zero123++: a single image to consistent multi-view dif- fusion base model. arXiv preprint arXiv:2310.15110, 2023. 3

    arXiv

  35. [43]

    Splat-mover: multi-stage, open-vocabulary robotic manipulation via editable gaussian splatting

    Olaolu Shorinwa, Johnathan Tucker, Aliyah Smith, Aiden Swann, Timothy Chen, Roya Firoozi, Monroe David Kennedy, and Mac Schwager. Splat-mover: multi-stage, open-vocabulary robotic manipulation via editable gaussian splatting. In 8th Annual Conference on Robot Learning ,

  36. [44]

    Resolution-robust large mask inpainting with fourier convolutions

    Roman Suvorov, Elizaveta Logacheva, Anton Mashikhin, Anastasia Remizova, Arsenii Ashukha, Aleksei Silvestrov, Naejin Kong, Harshith Goka, Kiwoong Park, and Victor Lempitsky. Resolution-robust large mask inpainting with fourier convolutions. InProceedings of the IEEE/CVF winter...

    2022

  37. [45]

    Habitat 2.0: Training home assistants to rearrange their habitat

    Andrew Szot, Alex Clegg, Eric Undersander, Erik Wijmans, Yili Zhao, John Turner, Noah Maestre, Mustafa Mukadam, Devendra Chaplot, Oleksandr Maksymets, Aaron Gokaslan, Vladimir V ondrus, Sameer Dharur, Franziska Meier, Woj- ciech Galuba, Angel Chang, Zsolt Kira, Vladlen Koltun,...

  38. [46]

    Open- mask3d: Open-vocabulary 3d instance segmentation

    Ayc ¸a Takmaz, Elisabetta Fedele, Robert W Sumner, Marc Pollefeys, Federico Tombari, and Francis Engelmann. Open- mask3d: Open-vocabulary 3d instance segmentation. arXiv preprint arXiv:2306.13631, 2023. 1

    2023 arXiv

  39. [47]

    Neus: Learning neural implicit surfaces by volume rendering for multi-view reconstruction

    Peng Wang, Lingjie Liu, Yuan Liu, Christian Theobalt, Taku Komura, and Wenping Wang. Neus: Learning neural implicit surfaces by volume rendering for multi-view reconstruction. arXiv preprint arXiv:2106.10689, 2021. 1, 7

    2021 arXiv

  40. [48]

    Nerfiller: Completing scenes via generative 3d inpainting

    Ethan Weber, Aleksander Holynski, Varun Jampani, Saurabh Saxena, Noah Snavely, Abhishek Kar, and Angjoo Kanazawa. Nerfiller: Completing scenes via generative 3d inpainting. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 20731– 20741...

    2024

  41. [49]

    Object- compositional neural implicit surfaces

    Qianyi Wu, Xian Liu, Yuedong Chen, Kejie Li, Chuanxia Zheng, Jianfei Cai, and Jianmin Zheng. Object- compositional neural implicit surfaces. In European Con- ference on Computer Vision, pages 197–213. Springer, 2022. 2

    2022

  42. [50]

    Objectsdf++: Improved object-compositional neural implicit surfaces

    Qianyi Wu, Kaisiyuan Wang, Kejie Li, Jianmin Zheng, and Jianfei Cai. Objectsdf++: Improved object-compositional neural implicit surfaces. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 21764– 21774, 2023. 2

    2023

  43. [51]

    Opengaussian: Towards point-level 3d gaussian-based open vocabulary understanding

    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. arXiv preprint arXiv:2406.02058, 2024. 3, 4

    2024 arXiv

  44. [52]

    Sparp: Fast 3d object reconstruction and pose estimation from sparse views

    Chao Xu, Ang Li, Linghao Chen, Yulin Liu, Ruoxi Shi, Hao Su, and Minghua Liu. Sparp: Fast 3d object reconstruction and pose estimation from sparse views. InEuropean Confer- ence on Computer Vision, pages 143–163. Springer, 2025. 4, 7

    2025

  45. [53]

    Instantmesh: Efficient 3d mesh generation from a single image with sparse-view large reconstruction models

    Jiale Xu, Weihao Cheng, Yiming Gao, Xintao Wang, Shenghua Gao, and Ying Shan. Instantmesh: Efficient 3d mesh generation from a single image with sparse-view large reconstruction models. arXiv preprint arXiv:2404.07191 ,

    arXiv

  46. [54]

    Maskclus- tering: View consensus based mask graph clustering for open-vocabulary 3d instance segmentation

    Mi Yan, Jiazhao Zhang, Yan Zhu, and He Wang. Maskclus- tering: View consensus based mask graph clustering for open-vocabulary 3d instance segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pat- tern Recognition, pages 28274–28284, 2024. 4

    2024

  47. [55]

    Street gaussians for modeling dynamic ur- ban scenes

    Yunzhi Yan, Haotong Lin, Chenxu Zhou, Weijie Wang, Haiyang Sun, Kun Zhan, Xianpeng Lang, Xiaowei Zhou, and Sida Peng. Street gaussians for modeling dynamic ur- ban scenes. arXiv preprint arXiv:2401.01339, 2024. 1

    2024 arXiv

  48. [56]

    Learning object-compositional neural radiance field for ed- itable scene rendering

    Bangbang Yang, Yinda Zhang, Yinghao Xu, Yijin Li, Han Zhou, Hujun Bao, Guofeng Zhang, and Zhaopeng Cui. Learning object-compositional neural radiance field for ed- itable scene rendering. In International Conference on Com- puter Vision (ICCV), 2021. 2

    2021

  49. [57]

    Sam3d: Segment anything in 3d scenes.arXiv preprint arXiv:2306.03908, 2023

    Yunhan Yang, Xiaoyang Wu, Tong He, Hengshuang Zhao, and Xihui Liu. Sam3d: Segment anything in 3d scenes.arXiv preprint arXiv:2306.03908, 2023. 4

    2023 arXiv

  50. [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. In European Conference on Computer Vision , pages 162–

  51. [59]

    Sai3d: Segment any instance in 3d scenes

    Yingda Yin, Yuzheng Liu, Yang Xiao, Daniel Cohen-Or, Jingwei Huang, and Baoquan Chen. Sai3d: Segment any instance in 3d scenes. In Proceedings of the IEEE/CVF Con- ference on Computer Vision and Pattern Recognition, pages 3292–3302, 2024. 4

    2024

  52. [60]

    Omniseg3d: Omniversal 3d segmentation via hierarchical contrastive learning

    Haiyang Ying, Yixuan Yin, Jinzhi Zhang, Fan Wang, Tao Yu, Ruqi Huang, and Lu Fang. Omniseg3d: Omniversal 3d segmentation via hierarchical contrastive learning. In Pro- ceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 20612–20622, 2024. 3, 4

    2024

  53. [61]

    Clay: A controllable large-scale generative model for creat- ing high-quality 3d assets

    Longwen Zhang, Ziyu Wang, Qixuan Zhang, Qiwei Qiu, Anqi Pang, Haoran Jiang, Wei Yang, Lan Xu, and Jingyi Yu. Clay: A controllable large-scale generative model for creat- ing high-quality 3d assets. ACM Transactions on Graphics (TOG), 43(4):1–20, 2024. 1, 2

    2024

  54. [62]

    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 Compu...

    2024

  55. [63]

    Drivinggaussian: Composite gaussian splatting for surrounding dynamic au- tonomous driving scenes

    Xiaoyu Zhou, Zhiwei Lin, Xiaojun Shan, Yongtao Wang, Deqing Sun, and Ming-Hsuan Yang. Drivinggaussian: Composite gaussian splatting for surrounding dynamic au- tonomous driving scenes. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , pages...

    2024

  56. [179]

    Springer, 2025. 3, 5, 6

    2025

Pith tools

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