pith. machine review for the scientific record. sign in

arxiv: 2605.03463 · v1 · submitted 2026-05-05 · 💻 cs.CV

Recognition: unknown

First Shape, Then Meaning: Efficient Geometry and Semantics Learning for Indoor Reconstruction

Remi Chierchia , L\'eo Lebrat , David Ahmedt-Aristizabal , Olivier Salvado , Clinton Fookes , Rodrigo Santa Cruz
Authors on Pith no claims yet

Pith reviewed 2026-05-07 17:54 UTC · model grok-4.3

classification 💻 cs.CV
keywords geometrysemanticfstmindoorlearningmulti-sdfreconstructionscene
0
0 comments X

The pith

FSTM improves indoor reconstruction by training geometry first without semantic supervision, then adding semantics, achieving 2.3x faster training and higher object surface recall than joint optimization.

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

Creating accurate 3D models of indoor spaces like rooms and offices is useful for many technologies, from self-driving robots to virtual reality. Recent methods use neural networks to represent the shape of surfaces with something called signed distance functions, which tell how close any point in space is to the nearest surface. To make these models more useful, researchers also want to know what each part of the scene represents, such as whether a surface is a wall, floor, or piece of furniture. Previous approaches tried to learn both the shape and these meanings at the same time, but this joint training is slow and doesn't scale well to large scenes. The new method, called FSTM, changes the order of learning. It first focuses only on getting the geometry right by using color images and some basic geometric information. After the shape is learned, it then trains the network to assign semantic labels to different parts of the scene. Experiments show that this separation helps a lot. The training happens 2.3 times faster on one dataset, and the models are better at handling real-world data that has imperfections. Importantly, it can recover the surfaces of more objects in the scene than methods that learn everything together. By keeping the approach simple without adding special modules, FSTM demonstrates that careful ordering of the learning steps can lead to better and more efficient results in 3D reconstruction.

Core claim

By first optimising geometry without semantic supervision, we observe substantial improvements compared to the standard joint optimisation.

Load-bearing premise

That a geometry warm-up phase using only RGB inputs and geometric cues provides a sufficiently accurate and unbiased base for subsequent semantic field estimation without missing or distorting semantic information.

Figures

Figures reproduced from arXiv: 2605.03463 by Clinton Fookes, David Ahmedt-Aristizabal, L\'eo Lebrat, Olivier Salvado, Remi Chierchia, Rodrigo Santa Cruz.

Figure 1
Figure 1. Figure 1: Qualitative comparison of 3D object segmentation for the ScanNet++ dataset of scene view at source ↗
Figure 2
Figure 2. Figure 2: The optimisation of multi-SDF architectures degrades view at source ↗
Figure 3
Figure 3. Figure 3: FSTM framework. For clarity, the dependence on input variables is omitted. Panels a) and b) illustrate the two architectures: view at source ↗
Figure 4
Figure 4. Figure 4: Qualitative evaluation of an object-level reconstruction on the Replica dataset. Our two-step method more accurately view at source ↗
Figure 5
Figure 5. Figure 5: Qualitative comparison of 3D object segmentation for scene view at source ↗
Figure 6
Figure 6. Figure 6: Convergence behaviour across the 8 Replica scenes from view at source ↗
Figure 7
Figure 7. Figure 7: Colour, depth and normal map novel-view renderings for Replica scene view at source ↗
Figure 8
Figure 8. Figure 8: Novel-view 2D segmentation renderings for Replica view at source ↗
Figure 9
Figure 9. Figure 9: Evaluation under sparse semantic supervision across the view at source ↗
read the original abstract

Neural Surface Reconstruction has become a standard methodology for indoor 3D reconstruction, with Signed Distance Functions (SDFs) proving particularly effective for representing scene geometry. A variety of applications require a detailed understanding of the scene context, driving the need for object-level semantic signals. While recent methods successfully integrate semantic labels, they often inherit the slow training time and limited scalability of multi-SDF learning. In this paper, we introduce FSTM, a unified approach for learning geometry and semantics through a two-step process: a geometry warm-up using RGB inputs and geometric cues, followed by semantic field estimation. By first optimising geometry without semantic supervision, we observe substantial improvements compared to the standard joint optimisation. Rather than relying on specialised modules or complex multi-SDF designs, FSTM shows that a streamlined formulation is sufficient to achieve strong geometric and semantic reconstructions. Experiments on both synthetic and real-world indoor datasets show that our method outperforms multi-SDF approaches. It trains 2.3x faster on Replica, improves robustness to real-world imperfections on ScanNet++, and achieves higher recall by recovering the surfaces of more objects in the scene. The code will be made available at https://remichierchia.github.io/FSTM.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Referee Report

3 major / 2 minor

Summary. The paper introduces FSTM, a two-phase neural surface reconstruction method for indoor scenes: a geometry warm-up phase using only RGB inputs and geometric cues (without semantic supervision), followed by semantic field estimation. It claims this separation yields substantial gains over standard joint optimization of geometry and semantics, including 2.3x faster training on Replica, improved robustness to real-world imperfections on ScanNet++, and higher recall by recovering surfaces of more objects, all without specialized modules or multi-SDF designs.

Significance. If the reported gains hold under controlled comparisons, the work would demonstrate that a simple ordering of optimization objectives can outperform more complex joint or multi-SDF formulations, potentially improving training efficiency and scalability for semantic indoor reconstruction. The promise of releasing code is a positive factor for reproducibility.

major comments (3)
  1. [Abstract] Abstract: The central claim of 'substantial improvements' (2.3x speedup on Replica, higher robustness on ScanNet++, higher object recall) is asserted without any reported metrics, baselines, ablation controls, or quantitative tables, preventing verification that the two-phase ordering itself drives the gains rather than training schedule or total iterations.
  2. [Method] Method (two-step process description): It is not specified whether geometry parameters are frozen after the warm-up phase or continue to receive gradients during semantic field estimation; if the latter, the advantage may reduce to a loss-weighting schedule that could be replicated inside a single joint optimization loop, undermining the claim that explicit separation is required.
  3. [Experiments] Experiments: No details are provided on how the 'standard joint optimisation' baseline is implemented (e.g., loss weights, iteration counts, or whether it matches total compute of the two-phase procedure), making it impossible to isolate the effect of the geometry-first ordering from confounding factors such as curriculum effects.
minor comments (2)
  1. [Abstract] The abstract introduces the acronym FSTM without expansion; this should be spelled out on first use.
  2. [Related Work] The claim that the method 'avoids specialised modules' would benefit from a brief comparison table listing the architectural components of the cited multi-SDF baselines.

Simulated Author's Rebuttal

3 responses · 0 unresolved

We thank the referee for their thoughtful review and constructive suggestions. We address each major comment below and outline the revisions to be incorporated in the updated manuscript.

read point-by-point responses

  1. Referee: [Abstract] Abstract: The central claim of 'substantial improvements' (2.3x speedup on Replica, higher robustness on ScanNet++, higher object recall) is asserted without any reported metrics, baselines, ablation controls, or quantitative tables, preventing verification that the two-phase ordering itself drives the gains rather than training schedule or total iterations.

    Authors: While the abstract summarizes the main results, the full quantitative evidence, including metrics, baselines, and ablations, is presented in the Experiments section with supporting tables and figures. To better highlight the claims, we will revise the abstract to incorporate specific numerical values for the speedup and recall improvements. This will make the central claims more verifiable directly from the abstract while referring readers to the detailed comparisons that isolate the effect of the two-phase ordering. revision: partial

  2. Referee: [Method] Method (two-step process description): It is not specified whether geometry parameters are frozen after the warm-up phase or continue to receive gradients during semantic field estimation; if the latter, the advantage may reduce to a loss-weighting schedule that could be replicated inside a single joint optimization loop, undermining the claim that explicit separation is required.

    Authors: The geometry parameters continue to be updated during the semantic field estimation phase. However, the two-phase approach differs from a simple loss-weighting schedule in a joint optimization because the warm-up phase allows the network to first learn a high-quality geometric representation without the influence of semantic losses, which can otherwise lead to suboptimal geometry. Our experiments include comparisons to joint optimization with different weighting schemes, demonstrating the superiority of the explicit separation. We will explicitly state this in the Method section and include a diagram or pseudocode clarifying the optimization flow. revision: yes

  3. Referee: [Experiments] Experiments: No details are provided on how the 'standard joint optimisation' baseline is implemented (e.g., loss weights, iteration counts, or whether it matches total compute of the two-phase procedure), making it impossible to isolate the effect of the geometry-first ordering from confounding factors such as curriculum effects.

    Authors: We agree that additional details are required for reproducibility and fair comparison. In the revised manuscript, we will describe the implementation of the joint optimization baseline, specifying the loss weights used, the total iteration count (matched to the two-phase method's total compute), and include ablations that control for curriculum effects by varying the introduction of semantic losses in a single loop. This will confirm that the gains are due to the geometry-first ordering. revision: yes

Axiom & Free-Parameter Ledger

0 free parameters · 0 axioms · 0 invented entities

Abstract-only review provides no specific parameters, axioms, or new entities; the contribution is a training procedure rather than a new mathematical formulation.

pith-pipeline@v0.9.0 · 5536 in / 1158 out tokens · 127500 ms · 2026-05-07T17:54:00.102108+00:00 · methodology

Review history (3 revisions) →

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

62 extracted references · 9 canonical work pages · 1 internal anchor

  1. [1]

    Nerf-slam: Real- time dense monocular slam with neural radiance fields,

    A. Rosinol, J. J. Leonard, and L. Carlone, “Nerf-slam: Real- time dense monocular slam with neural radiance fields,” in IEEE/RSJ Int. Conf. Intell. Robots Syst. (IROS). IEEE, 2023, pp. 3437–3444

    2023

  2. [2]

    Vision-only robot navi- gation in a neural radiance world,

    M. Adamkiewicz, T. Chen, A. Caccavale, R. Gardner, P. Cul- bertson, J. Bohg, and M. Schwager, “Vision-only robot navi- gation in a neural radiance world,”IEEE Robot. Autom. Lett., vol. 7, no. 2, pp. 4606–4613, 2022

    2022

  3. [3]

    Magic3d: High-resolution text-to-3d content creation,

    C.-H. Lin, J. Gao, L. Tang, T. Takikawa, X. Zeng, X. Huang, K. Kreis, S. Fidler, M.-Y . Liu, and T.-Y . Lin, “Magic3d: High-resolution text-to-3d content creation,” inProc. IEEE/CVF Conf. Comput. Vis. Pattern Recognit. (CVPR), 2023, pp. 300–309

    2023

  4. [4]

    Endosurf: Neural surface reconstruction of deformable tissues with stereo endoscope videos,

    R. Zha, X. Cheng, H. Li, M. Harandi, and Z. Ge, “Endosurf: Neural surface reconstruction of deformable tissues with stereo endoscope videos,” inInt. Conf. Med. Image Comput. Comput.-Assist. Interv. (MICCAI). Springer, 2023, pp. 13– 23

    2023

  5. [5]

    Salve: A 3d reconstruction benchmark of wounds from consumer-grade videos,

    R. Chierchia, L. Lebrat, D. Ahmedt-Aristizabal, O. Salvado, C. Fookes, and R. S. Cruz, “Salve: A 3d reconstruction benchmark of wounds from consumer-grade videos,” in IEEE/CVF Winter Conf. Appl. Comput. Vis. (WACV), 2025, pp. 4205–4214

    2025

  6. [6]

    arXiv preprint arXiv:2106.10689 , year=

    P. Wang, L. Liu, Y . Liu, C. Theobalt, T. Komura, and W. Wang, “Neus: Learning neural implicit surfaces by vol- ume rendering for multi-view reconstruction,”arXiv preprint arXiv:2106.10689, 2021

    work page arXiv 2021

  7. [7]

    Multiview neural surface reconstruc- tion by disentangling geometry and appearance,

    L. Yariv, Y . Kasten, D. Moran, M. Galun, M. Atzmon, B. Ro- nen, and Y . Lipman, “Multiview neural surface reconstruc- tion by disentangling geometry and appearance,”Adv. Neural Inf. Process. Syst., vol. 33, pp. 2492–2502, 2020

    2020

  8. [8]

    Sdfstudio: A unified framework for surface reconstruction,

    Z. Yu, A. Chen, B. Antic, S. Peng, A. Bhattacharyya, M. Niemeyer, S. Tang, T. Sattler, and A. Geiger, “Sdfstudio: A unified framework for surface reconstruction,” 2022. [Online]. Available: https://github.com/autonomousvision/ sdfstudio

    2022

  9. [9]

    Objectsdf++: Improved object-compositional neural implicit surfaces,

    Q. Wu, K. Wang, K. Li, J. Zheng, and J. Cai, “Objectsdf++: Improved object-compositional neural implicit surfaces,” in Proc. IEEE/CVF Int. Conf. Comput. Vis. (ICCV), October 2023, pp. 21 764–21 774

    2023

  10. [10]

    Rico: Regularizing the unobservable for indoor compositional re- construction,

    Z. Li, X. Lyu, Y . Ding, M. Wang, Y . Liao, and Y . Liu, “Rico: Regularizing the unobservable for indoor compositional re- construction,” inProc. IEEE/CVF Int. Conf. Comput. Vis. (ICCV), October 2023, pp. 17 761–17 771

    2023

  11. [11]

    Neural radiance fields for the real world: A survey.arXiv preprint arXiv:2501.13104, 2025

    W. Xiao, R. Chierchia, R. S. Cruz, X. Li, D. Ahmedt- Aristizabal, O. Salvado, C. Fookes, and L. Lebrat, “Neural radiance fields for the real world: A survey,”arXiv preprint arXiv:2501.13104, 2025. 11

    work page arXiv 2025

  12. [12]

    Neuralangelo: High-fidelity neural surface reconstruction,

    Z. Li, T. Müller, A. Evans, R. H. Taylor, M. Unberath, M.- Y . Liu, and C.-H. Lin, “Neuralangelo: High-fidelity neural surface reconstruction,” inProc. IEEE/CVF Conf. Comput. Vis. Pattern Recognit. (CVPR), June 2023, pp. 8456–8465

    2023

  13. [13]

    Geo-neus: Geometry- consistent neural implicit surfaces learning for multi-view reconstruction,

    Q. Fu, Q. Xu, Y . S. Ong, and W. Tao, “Geo-neus: Geometry- consistent neural implicit surfaces learning for multi-view reconstruction,” inAdv. Neural Inf. Process. Syst., vol. 35, 2022, pp. 3403–3416

    2022

  14. [14]

    Learning signed dis- tance field for multi-view surface reconstruction,

    J. Zhang, Y . Yao, and L. Quan, “Learning signed dis- tance field for multi-view surface reconstruction,” inProc. IEEE/CVF Int. Conf. Comput. Vis. (ICCV), October 2021, pp. 6525–6534

    2021

  15. [15]

    Improving neural implicit surfaces geometry with patch warping,

    F. Darmon, B. Bascle, J.-C. Devaux, P. Monasse, and M. Aubry, “Improving neural implicit surfaces geometry with patch warping,” inProc. IEEE/CVF Conf. Comput. Vis. Pattern Recognit. (CVPR), June 2022, pp. 6260–6269

    2022

  16. [16]

    Critical regularizations for neural surface recon- struction in the wild,

    J. Zhang, Y . Yao, S. Li, T. Fang, D. McKinnon, Y . Tsin, and L. Quan, “Critical regularizations for neural surface recon- struction in the wild,” inProc. IEEE/CVF Conf. Comput. Vis. Pattern Recognit. (CVPR), June 2022, pp. 6270–6279

    2022

  17. [17]

    Regnerf: Regularizing neural ra- diance fields for view synthesis from sparse inputs,

    M. Niemeyer, J. T. Barron, B. Mildenhall, M. S. M. Sajjadi, A. Geiger, and N. Radwan, “Regnerf: Regularizing neural ra- diance fields for view synthesis from sparse inputs,” inProc. IEEE/CVF Conf. Comput. Vis. Pattern Recognit. (CVPR), June 2022, pp. 5480–5490

    2022

  18. [18]

    Unisurf: Unifying neural implicit surfaces and radiance fields for multi-view reconstruction,

    M. Oechsle, S. Peng, and A. Geiger, “Unisurf: Unifying neural implicit surfaces and radiance fields for multi-view reconstruction,” inProc. IEEE/CVF Int. Conf. Comput. Vis. (ICCV), October 2021, pp. 5589–5599

    2021

  19. [19]

    Monosdf: Exploring monocular geometric cues for neural implicit surface reconstruction,

    Z. Yu, S. Peng, M. Niemeyer, T. Sattler, and A. Geiger, “Monosdf: Exploring monocular geometric cues for neural implicit surface reconstruction,” inAdv. Neural Inf. Process. Syst., vol. 35, 2022, pp. 25 018–25 032

    2022

  20. [20]

    Neuris: Neural reconstruction of indoor scenes using normal priors,

    J. Wang, P. Wang, X. Long, C. Theobalt, T. Komura, L. Liu, and W. Wang, “Neuris: Neural reconstruction of indoor scenes using normal priors,” inEur. Conf. Comput. Vis. (ECCV). Springer, 2022, pp. 139–155

    2022

  21. [21]

    Sni-slam: Semantic neural implicit slam,

    S. Zhu, G. Wang, H. Blum, J. Liu, L. Song, M. Pollefeys, and H. Wang, “Sni-slam: Semantic neural implicit slam,” inProc. IEEE/CVF Conf. Comput. Vis. Pattern Recognit. (CVPR), 2024, pp. 21 167–21 177

    2024

  22. [22]

    Surroundsdf: Implicit 3d scene understanding based on signed distance field,

    L. Liu, B. Wang, H. Xie, D. Liu, L. Liu, Z. Tian, K. Yang, and B. Wang, “Surroundsdf: Implicit 3d scene understanding based on signed distance field,” inProc. IEEE/CVF Conf. Comput. Vis. Pattern Recognit. (CVPR), 2024, pp. 21 614– 21 623

    2024

  23. [23]

    Ov-nerf: Open- vocabulary neural radiance fields with vision and language foundation models for 3d semantic understanding,

    G. Liao, K. Zhou, Z. Bao, K. Liu, and Q. Li, “Ov-nerf: Open- vocabulary neural radiance fields with vision and language foundation models for 3d semantic understanding,”IEEE Trans. Circuits Syst. Video Technol., vol. 34, no. 12, pp. 12 923–12 936, 2024

    2024

  24. [24]

    Decompos- ing nerf for editing via feature field distillation,

    S. Kobayashi, E. Matsumoto, and V . Sitzmann, “Decompos- ing nerf for editing via feature field distillation,”Adv. Neural Inf. Process. Syst., vol. 35, pp. 23 311–23 330, 2022

    2022

  25. [25]

    Segment anything in 3d with nerfs,

    J. Cen, Z. Zhou, J. Fang, c. yang, W. Shen, L. Xie, D. Jiang, X. Zhang, and Q. Tian, “Segment anything in 3d with nerfs,” inAdv. Neural Inf. Process. Syst., A. Oh, T. Naumann, A. Globerson, K. Saenko, M. Hardt, and S. Levine, Eds., vol. 36. Curran Associates, Inc., 2023, pp. 25 971–25 990

    2023

  26. [26]

    Navinerf: Nerf-based 3d representation disentan- glement by latent semantic navigation,

    B. Xie, B. Li, Z. Zhang, J. Dong, X. Jin, J. Yang, and W. Zeng, “Navinerf: Nerf-based 3d representation disentan- glement by latent semantic navigation,” inProc. IEEE/CVF Int. Conf. Comput. Vis. (ICCV), October 2023, pp. 17 992– 18 002

    2023

  27. [27]

    Repaint-nerf: Nerf editting via semantic masks and diffusion models,

    X. Zhou, Y . He, F. R. Yu, J. Li, and Y . Li, “Repaint-nerf: Nerf editting via semantic masks and diffusion models,” in Proc. Int. Joint Conf. Artif. Intell., ser. IJCAI ’23, 2023

    2023

  28. [28]

    Instruct-nerf2nerf: Editing 3d scenes with instructions,

    A. Haque, M. Tancik, A. A. Efros, A. Holynski, and A. Kanazawa, “Instruct-nerf2nerf: Editing 3d scenes with instructions,” inProc. IEEE/CVF Int. Conf. Comput. Vis. (ICCV), October 2023, pp. 19 740–19 750

    2023

  29. [29]

    In- place scene labelling and understanding with implicit scene representation,

    S. Zhi, T. Laidlow, S. Leutenegger, and A. J. Davison, “In- place scene labelling and understanding with implicit scene representation,” inProc. IEEE/CVF Int. Conf. Comput. Vis. (ICCV), October 2021, pp. 15 838–15 847

    2021

  30. [30]

    Panoptic nerf: 3d-to-2d label transfer for panoptic urban scene segmentation,

    X. Fu, S. Zhang, T. Chen, Y . Lu, L. Zhu, X. Zhou, A. Geiger, and Y . Liao, “Panoptic nerf: 3d-to-2d label transfer for panoptic urban scene segmentation,” inInt. Conf. 3D Vis. (3DV), 2022, pp. 1–11

    2022

  31. [31]

    Panoptic lifting for 3d scene understanding with neural fields,

    Y . Siddiqui, L. Porzi, S. R. Bulò, N. Müller, M. Nießner, A. Dai, and P. Kontschieder, “Panoptic lifting for 3d scene understanding with neural fields,” inProc. IEEE/CVF Conf. Comput. Vis. Pattern Recognit. (CVPR), June 2023, pp. 9043– 9052

    2023

  32. [32]

    Clusteringsdf: Self-organized neural implicit surfaces for 3d decomposi- tion,

    T. Wu, C. Zheng, Q. Wu, and T.-J. Cham, “Clusteringsdf: Self-organized neural implicit surfaces for 3d decomposi- tion,” inEur. Conf. Comput. Vis. (ECCV). Springer, 2024, pp. 255–272

    2024

  33. [33]

    Putting nerf on a diet: Semantically consistent few-shot view synthesis,

    A. Jain, M. Tancik, and P. Abbeel, “Putting nerf on a diet: Semantically consistent few-shot view synthesis,” inProc. IEEE/CVF Int. Conf. Comput. Vis. (ICCV), October 2021, pp. 5885–5894

    2021

  34. [34]

    Sinnerf: Training neural radiance fields on complex scenes from a single image,

    D. Xu, Y . Jiang, P. Wang, Z. Fan, H. Shi, and Z. Wang, “Sinnerf: Training neural radiance fields on complex scenes from a single image,” inEur. Conf. Comput. Vis. (ECCV). Springer, 2022, pp. 736–753. 12

    2022

  35. [35]

    Nesf: Neural semantic fields for generalizable semantic segmenta- tion of 3d scenes,

    S. V ora, N. Radwan, K. Greff, H. Meyer, K. Genova, M. S. Sajjadi, E. Pot, A. Tagliasacchi, and D. Duckworth, “Nesf: Neural semantic fields for generalizable semantic segmenta- tion of 3d scenes,”arXiv preprint arXiv:2111.13260, 2021

    work page arXiv 2021

  36. [36]

    Semantic ray: Learning a generalizable semantic field with cross- reprojection attention,

    F. Liu, C. Zhang, Y . Zheng, and Y . Duan, “Semantic ray: Learning a generalizable semantic field with cross- reprojection attention,” inProc. IEEE/CVF Conf. Com- put. Vis. Pattern Recognit. (CVPR), June 2023, pp. 17 386– 17 396

    2023

  37. [37]

    Neural 3d scene reconstruction with the manhattan-world assumption,

    H. Guo, S. Peng, H. Lin, Q. Wang, G. Zhang, H. Bao, and X. Zhou, “Neural 3d scene reconstruction with the manhattan-world assumption,” inProc. IEEE/CVF Conf. Comput. Vis. Pattern Recognit. (CVPR), June 2022, pp. 5511– 5520

    2022

  38. [38]

    H2o-sdf: two-phase learning for 3d indoor reconstruction us- ing object surface fields,

    M. Park, M. Do, Y . Shin, J. Yoo, J. Hong, J. Kim, and C. Lee, “H2o-sdf: two-phase learning for 3d indoor reconstruction us- ing object surface fields,”arXiv preprint arXiv:2402.08138, 2024

    work page arXiv 2024

  39. [39]

    Object-compositional neural implicit surfaces,

    Q. Wu, X. Liu, Y . Chen, K. Li, C. Zheng, J. Cai, and J. Zheng, “Object-compositional neural implicit surfaces,” inEur. Conf. Comput. Vis. (ECCV). Springer, 2022, pp. 197–213

    2022

  40. [40]

    Total- decom: Decomposed 3d scene reconstruction with minimal interaction,

    X. Lyu, C. Chang, P. Dai, Y .-T. Sun, and X. Qi, “Total- decom: Decomposed 3d scene reconstruction with minimal interaction,” inProc. IEEE/CVF Conf. Comput. Vis. Pattern Recognit. (CVPR), June 2024, pp. 20 860–20 869

    2024

  41. [41]

    Phyrecon: Physically plausible neural scene reconstruction,

    J. Ni, Y . Chen, B. Jing, N. Jiang, B. Wang, B. Dai, P. Li, Y . Zhu, S.-C. Zhu, and S. Huang, “Phyrecon: Physically plausible neural scene reconstruction,” inAdv. Neural Inf. Process. Syst., vol. 37. Curran Associates, Inc., 2024, pp. 25 747–25 780

    2024

  42. [42]

    Decompositional neural scene reconstruction with generative diffusion prior,

    J. Ni, Y . Liu, R. Lu, Z. Zhou, S.-C. Zhu, Y . Chen, and S. Huang, “Decompositional neural scene reconstruction with generative diffusion prior,” inProc. IEEE/CVF Conf. Comput. Vis. Pattern Recognit. (CVPR), June 2025, pp. 6022– 6033

    2025

  43. [43]

    3d gaussian splatting for real-time radiance field rendering

    B. Kerbl, G. Kopanas, T. Leimkühler, G. Drettakiset al., “3d gaussian splatting for real-time radiance field rendering.” ACM Trans. Graph., vol. 42, no. 4, pp. 139–1, 2023

    2023

  44. [44]

    Gaussian group- ing: Segment and edit anything in 3d scenes,

    M. Ye, M. Danelljan, F. Yu, and L. Ke, “Gaussian group- ing: Segment and edit anything in 3d scenes,” inEur. Conf. Comput. Vis. (ECCV). Springer, 2024, pp. 162–179

    2024

  45. [45]

    Plgs: Robust panop- tic lifting with 3d gaussian splatting,

    Y . Wang, X. Wei, M. Lu, and G. Kang, “Plgs: Robust panop- tic lifting with 3d gaussian splatting,”IEEE Trans. Image Process., vol. 34, pp. 3377–3388, 2025

    2025

  46. [46]

    Panopticsplatting: End-to-end panoptic gaus- sian splatting,

    Y . Xie, X. Yu, C. Jiang, S. Mao, S. Zhou, R. Fan, R. Xiong, and Y . Wang, “Panopticsplatting: End-to-end panoptic gaus- sian splatting,” inIEEE/RSJ Int. Conf. Intell. Robots Syst. (IROS). IEEE, 2025, pp. 4067–4074

    2025

  47. [47]

    Objectgs: Object-aware scene reconstruction and scene understanding via gaussian splatting,

    R. Zhu, M. Yu, L. Xu, L. Jiang, Y . Li, T. Zhang, J. Pang, and B. Dai, “Objectgs: Object-aware scene reconstruction and scene understanding via gaussian splatting,” inProc. IEEE/CVF Int. Conf. Comput. Vis. (ICCV), 2025, pp. 8350– 8360

    2025

  48. [48]

    Learning object-compositional neural radi- ance field for editable scene rendering,

    B. Yang, Y . Zhang, Y . Xu, Y . Li, H. Zhou, H. Bao, G. Zhang, and Z. Cui, “Learning object-compositional neural radi- ance field for editable scene rendering,” inProc. IEEE/CVF Int. Conf. Comput. Vis. (ICCV), October 2021, pp. 13 779– 13 788

    2021

  49. [49]

    Deepsdf: Learning continuous signed distance func- tions for shape representation,

    J. J. Park, P. Florence, J. Straub, R. Newcombe, and S. Love- grove, “Deepsdf: Learning continuous signed distance func- tions for shape representation,” inProc. IEEE/CVF Conf. Comput. Vis. Pattern Recognit. (CVPR), June 2019

    2019

  50. [50]

    Implicit geometric regularization for learning shapes,

    A. Gropp, L. Yariv, N. Haim, M. Atzmon, and Y . Lipman, “Implicit geometric regularization for learning shapes,”arXiv preprint arXiv:2002.10099, 2020

    work page arXiv 2002

  51. [51]

    Sitzmann, J

    V . Sitzmann, J. N. P. Martel, A. W. Bergman, D. B. Lindell, and G. Wetzstein, “Implicit Neural Representations with Periodic Activation Functions,” 6 2020. [Online]. Available: http://arxiv.org/abs/2006.09661

    work page arXiv 2020

  52. [52]

    Nerf: representing scenes as neural radiance fields for view synthesis,

    B. Mildenhall, P. P. Srinivasan, M. Tancik, J. T. Barron, R. Ramamoorthi, and R. Ng, “Nerf: representing scenes as neural radiance fields for view synthesis,”Commun. ACM, vol. 65, no. 1, p. 99–106, Dec. 2021

    2021

  53. [53]

    V olume rendering of neural implicit surfaces,

    L. Yariv, J. Gu, Y . Kasten, and Y . Lipman, “V olume rendering of neural implicit surfaces,” inAdv. Neural Inf. Process. Syst., vol. 34, 2021, pp. 4805–4815

    2021

  54. [54]

    Omnidata: A scalable pipeline for making multi-task mid-level vision datasets from 3d scans,

    A. Eftekhar, A. Sax, J. Malik, and A. Zamir, “Omnidata: A scalable pipeline for making multi-task mid-level vision datasets from 3d scans,” inProc. IEEE/CVF Int. Conf. Com- put. Vis. (ICCV), October 2021, pp. 10 786–10 796

    2021

  55. [55]

    Pytorch: An imperative style, high-performance deep learning library,

    A. Paszke, S. Gross, F. Massa, A. Lerer, J. Bradbury, G. Chanan, T. Killeen, Z. Lin, N. Gimelshein, L. Antiga et al., “Pytorch: An imperative style, high-performance deep learning library,”Adv. Neural Inf. Process. Syst., vol. 32, 2019

    2019

  56. [56]

    Adam: A method for stochastic optimization,

    D. P. Kingma and J. Ba, “Adam: A method for stochastic optimization,” inInt. Conf. Learn. Represent. (ICLR), San Diego, CA, USA, May 7-9, 2015, Conference Track Proceed- ings, Y . Bengio and Y . LeCun, Eds., 2015

    2015

  57. [57]

    The Replica Dataset: A Digital Replica of Indoor Spaces

    J. Straub, T. Whelan, L. Ma, Y . Chen, E. Wijmans, S. Green, J. J. Engel, R. Mur-Artal, C. Ren, S. Vermaet al., “The Replica dataset: A digital replica of indoor spaces,”arXiv preprint arXiv:1906.05797, 2019

    work page internal anchor Pith review arXiv 1906

  58. [58]

    Scan- net++: A high-fidelity dataset of 3d indoor scenes,

    C. Yeshwanth, Y .-C. Liu, M. Nießner, and A. Dai, “Scan- net++: A high-fidelity dataset of 3d indoor scenes,” inProc. IEEE/CVF Int. Conf. Comput. Vis. (ICCV), 2023, pp. 12–22. 13

    2023

  59. [59]

    arXiv preprint arXiv:2411.16898 , year=

    K. Li, M. Niemeyer, Z. Chen, N. Navab, and F. Tombari, “Monogsdf: Exploring monocular geometric cues for gaus- sian splatting-guided implicit surface reconstruction,”arXiv preprint arXiv:2411.16898, 2024

    work page arXiv 2024

  60. [60]

    A lightweight approach to repairing digitized polygon meshes,

    M. Attene, “A lightweight approach to repairing digitized polygon meshes,”The visual computer, vol. 26, no. 11, pp. 1393–1406, 2010

    2010

  61. [61]

    Mv-dust3r+: Single-stage scene reconstruction from sparse views in 2 seconds,

    Z. Tang, Y . Fan, D. Wang, H. Xu, R. Ranjan, A. Schwing, and Z. Yan, “Mv-dust3r+: Single-stage scene reconstruction from sparse views in 2 seconds,” inProc. IEEE/CVF Conf. Comput. Vis. Pattern Recognit. (CVPR), 2025, pp. 5283– 5293

    2025

  62. [62]

    Fin3r: Fine-tuning feed-forward 3d reconstruction models via monocular knowl- edge distillation,

    W. Ren, H. Wang, X. Tan, and K. Han, “Fin3r: Fine-tuning feed-forward 3d reconstruction models via monocular knowl- edge distillation,”arXiv preprint arXiv:2511.22429, 2025. 14

    work page arXiv 2025