Pith. sign in

REVIEW 3 major objections 6 minor 44 references

Selecting non-keyframes that observe the most uncertain Gaussians and training them alongside keyframes makes online 3D Gaussian Splatting maps more complete and higher-quality than keyframe-only training, at real-time rates.

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

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

During online Gaussian splatting SLAM, training extra on non-keyframes that view the most uncertain Gaussians improves model completeness over keyframe-only training.

T0 review reviewed 2026-08-05 challenge →

load-bearing objection A genuinely new frame-selection idea for online 3DGS, but the completeness claim is running ahead of the evidence. the 3 major comments →

arxiv 2508.14014 v2 pith:EQ4D6XGH submitted 2025-08-19 cs.CV

Online 3D Gaussian Splatting Modeling with Novel View Selection

Byeonggwon Lee , Junkyu Park , Khang Truong Giang , Soohwan Song This is my paper
classification cs.CV
keywords 3D Gaussian Splattingonline mappingnovel view selectionGaussian uncertaintymulti-view stereoRGB-only SLAMreal-time reconstructionview synthesis
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

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

The reading

This paper tries to show that an online 3D Gaussian Splatting (3DGS) model built from RGB-only video can be made more complete by adaptively adding a small number of carefully chosen non-keyframes to the training set. The proposed selection signal is per-Gaussian uncertainty, computed from the shape and position-gradient statistics of each Gaussian, and views that project many high-uncertainty Gaussians are judged to carry the most information gain. Combined with a multi-view-stereo frontend that supplies consistent depth and global bundle adjustment with Gaussian deformation, the method reports better rendering and geometry than keyframe-only training, at real-time frame rates, on indoor benchmarks and on complex outdoor scenes.

Core claim

On the paper's own terms, the discovery is that reconstruction quality can be assessed online from the Gaussian field itself: Gaussians that are large or whose positions are still moving during optimization are marked uncertain, and a view's value is measured by how many of these uncertain Gaussians it observes. Selecting about 100 such non-keyframes and training them together with keyframes refines under-reconstructed regions from new viewpoints, yielding more complete models and higher rendering quality than training on keyframes alone. The paper further claims that integrating a learned multi-view stereo network with a disparity prior, plus continuous global bundle adjustment and Gaussian

What carries the argument

The load-bearing object is the Gaussian uncertainty map: each 3D Gaussian primitive receives an uncertainty value derived from its shape (large or elongated primitives are less trustworthy) and from the magnitude of its position gradients during optimization (still-moving primitives are not yet settled). A candidate non-keyframe's information gain is computed by projecting these uncertain Gaussians into the view and aggregating their uncertainty, so the selector can pick frames that observe the least-reconstructed regions. The companion machinery is the online MVS frontend, which estimates dense depth for keyframes and injects a disparity map into the MVS network, keeping the Gaussian traini

Load-bearing premise

The argument rests on the premise that high Gaussian uncertainty marks places where extra color-only training can actually reduce model error, rather than simply marking textureless surfaces, unreliable stereo depths, or inherently hard-to-render viewpoints.

What would settle it

Run the same 100-view selection ablation on a low-texture indoor scene and a complex outdoor scene with ground-truth geometry, measuring depth L1 and mesh completeness for uncertainty-selected views versus the same number of randomly selected views; if the uncertainty-selected model does not produce consistently better geometry, the uncertainty signal is tracking MVS noise or texturelessness instead of fixable incompleteness.

Watch this falsifier. Get emailed when new claim-graph text bears on it.

If this is right

  • Training on roughly 100 non-keyframes chosen by Gaussian uncertainty, together with keyframes, yields better rendering PSNR/SSIM/LPIPS and better depth L1 than keyframe-only training in the experiments reported.
  • Selected high-uncertainty views tend to be under-reconstructed ones: their rendered PSNR is markedly lower (about 19.6 versus 28.8) than low-uncertainty views, and their uncertainty values are roughly three times higher.
  • The pipeline stays real-time: with novel-view selection it averages about 9 FPS on the Replica office0 sequence, and every module's runtime stays below its invocation interval.
  • Injecting the DROID-SLAM disparity map into the MVS network improves both depth L1 (0.067 to 0.050) and rendering quality, so the depth prior is a measurable part of the gain.
  • Only 2K final-refinement iterations suffice to beat Splat-SLAM's 26K iterations on rendering quality in the reported comparison, indicating the online-trained field starts from a much better initialization.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • The uncertainty signal is really a next-best-view score in disguise: the same per-Gaussian uncertainty could steer an autonomous camera's motion toward unexplored regions, turning the offline frame-selection rule into an active reconstruction planner.
  • Because selected non-keyframes are trained with color only, the observed depth-L1 degradation at 150 added views is likely color-fitting winning over geometry; supplying depth for those views would be a direct test of whether completeness gains can be pushed further without sacrificing geometry.
  • The selection rule is agnostic to the depth source, so it should transfer to other MVS or depth-prediction backends; a straightforward check is running the same NVS logic on a system with a different frontend and comparing completeness.
  • Persistent high-uncertainty regions across many selected frames could double as a drift or gap detector for the SLAM system itself, flagging places where the trajectory or the map needs revisiting.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

3 major / 6 minor

Summary. The paper proposes an online 3D Gaussian Splatting (3DGS) mapping system for RGB-only image streams. The main novelty is a novel-view selection (NVS) module that estimates per-Gaussian uncertainty from shape and position-gradient statistics, scores incoming non-keyframes by predicted information gain, and adds the highest-uncertainty non-keyframes to the keyframe-based training set. The pipeline combines an MVS frontend (MVSFormer with DROID-SLAM disparity injection) and a parallel 3DGS backend with global bundle adjustment and Gaussian deformation. The paper reports evaluations on Replica, TUM-RGBD, ScanNet, Tanks & Temples, and an aerial dataset, with ablations on the number of selected views, selection criteria, disparity-map use, refinement iterations, and runtime.

Significance. If the central claim were fully established, the contribution would be useful: an online 3DGS mapper that adaptively chooses a small set of informative non-keyframes, improving reconstruction completeness while remaining real-time. The paper includes directed ablations that support the mechanism in a limited setting: uncertainty-based selection beats random and PSNR-based selection on Replica office0 (Table S8, PSNR 43.93 vs 43.29/43.35), disparity-map injection improves both rendering and depth (Table S7), and the system is evaluated on external benchmarks including outdoor scenes. However, the paper's headline claim is about geometric completeness, and that claim is not directly measured. The evidence actually reported is photometric quality and depth L1 on one ablation scene, plus a qualitative TartanAir figure. The runtime prose in §B.9 is also inconsistent with the table it cites. The core idea is plausible, but the current manuscript does not yet provide the geometric and multi-scene support needed to substantiate the completeness claim.

major comments (3)
  1. [§1, §B.5, Table S8, Fig. S4] The central claim is that NVS 'significantly enhances completeness,' but no geometric completeness metric is reported anywhere. Table S8, the only quantitative NVS ablation, reports PSNR/SSIM/LPIPS and depth L1 on a single Replica scene (office0); these are rendering and depth accuracy metrics, not completeness (e.g., F-score, accuracy/completeness on a reconstructed mesh/point cloud, or coverage of visible surfaces). Fig. S4 shows that high-uncertainty views have lower PSNR, which supports that the selector identifies hard-to-render viewpoints, but it does not show that training on those views improves geometric completeness. Please add a multi-scene evaluation with direct geometric completeness metrics, or revise the claim to 'improved rendering quality' rather than 'completeness.'
  2. [§B.4, Table S5] Table S5 is in tension with the mechanism claimed. At the operating point of 100 novel views, keyframe depth L1 improves (0.056→0.050) but every-5-frame depth L1 worsens (0.058→0.059); at 150 views, keyframe depth L1 also worsens (0.058 vs 0.056 baseline). The text states that 'the absence of depth information in novel views resulted in a reduction in reconstruction quality.' Since novel views are trained with color loss only, the evidence supports a trade-off rather than a geometric completeness gain. The manuscript should either provide depth-supervised novel-view training or a direct geometric evaluation that separates photometric overfitting from genuine 3D completion. Without this, the 100-view choice reads as a tuned compromise, not a validated mechanism.
  3. [§B.9, Table S9] The prose claims that the method 'delivers a significantly higher frame rate than MonoGS and Splat-SLAM,' but Table S9 lists Avg. FPS of 10.77 for Ours versus 17.57 for Splat-SLAM. Under the natural reading of the table, Splat-SLAM is about 1.6× faster. Please correct the claim or explain the measurement protocol (e.g., whether the comparison is on a different hardware/config); as written, the text contradicts the table.
minor comments (6)
  1. [§B.4, Table S5] The table caption does not define '0' in the 'Number of novel views' column; clarify that this is the keyframe-only baseline. Also, the FPS row shows a significant drop (12.07→9.18→8.54) that is discussed in prose but should be more prominent given the real-time claim.
  2. [§B.8, Table S8] The ablation is on a single scene and reports no variance or statistical significance. Please state whether the difference (e.g., depth L1 0.034 vs 0.042) is stable across runs or scenes; otherwise the comparison is anecdotal.
  3. [§B.9, Table S9] The column header 'Avg. FPS↑[GIB]↓' mixes FPS and GPU memory units; split into separate columns for readability. Also 'Table. S9' should be 'Table S9'.
  4. [Figure S5] The caption uses 'Scannet' while the text and references use 'ScanNet'; make capitalization consistent.
  5. [References] The reference [Redmon, 2016] appears in the bibliography but is not cited in the visible text; check whether it is needed or remove it.
  6. [§1] The introduction states that previous methods store depth only for keyframes, but the contribution over MVS-GS is not crisply stated. Consider a short paragraph explicitly listing the delta from [Lee et al., 2024] (GBA with Gaussian deformation, NVS, disparity fusion) to help readers position the work.

Circularity Check

1 steps flagged

Every-5-frame 'generalization' scores are computed after the selected non-keyframes are added to training, with no stated exclusion, so the central rendering/completeness gain is partly in-sample fitting.

specific steps
  1. fitted input called prediction [Appendix B.3, B.4 (Table S5)]
    "Every-5-frame rendering reflects the generalization performance of the 3D model, showing superior performance compared to [Zhang et al., 2024]. ... This highlights that our approach demonstrates excellent generalization capabilities even for unseen views. Increasing the number of novel views consistently improved rendering quality ... we selected 100 frames."

    Table S5 reports 'Every 5 Frames' metrics at 0, 100, and 150 novel views. The 100/150 views are the non-keyframes added to 3DGS training by NVS, so those images are fitted, not held out. The paper never states that selected NVS frames are excluded from the every-5-frame evaluation set. Therefore the reported PSNR/LPIPS improvements over the 0-view baseline can be partly a direct measure of fitting the very frames used for training, and the claim that these numbers demonstrate 'generalization ... even for unseen views' is not supported by the stated protocol. The 100-view operating point is also selected using this same contaminated metric, making the central completeness/rendering gain partially circular.

full rationale

The uncertainty-based view selector itself is not circular by definition: it is ablated against random selection and a PSNR-based selection baseline (Table S8), and the overall pipeline is evaluated on public indoor/outdoor benchmarks against external baselines. Self-citations to MVS-GS and Song et al. are used as the base system, but the NVS delta is independently tested, so those citations are not load-bearing. However, one evaluation step is genuinely circular: the paper's 'unseen view' generalization claim rests on every-5-frame metrics computed after the selected non-keyframes have been added to the training set, with no stated exclusion of those frames from evaluation. Table S5 itself shows that adding more such views improves rendering metrics while depth L1 degrades (e.g., keyframe Depth L1 goes 0.056 -> 0.050 -> 0.058), and the text attributes the gain to increased viewpoint diversity while acknowledging that 'the absence of depth information in novel views resulted in a reduction in reconstruction quality.' This means the headline completeness improvement is not established as a geometric gain; it may largely reflect photometric fitting to the selected views. Since the central claim is measured with the contaminated metric, the partial circularity score is 6 rather than 2.

Axiom & Free-Parameter Ledger

3 free parameters · 5 axioms · 0 invented entities

No new physical entities are introduced. The paper's proposed constructs are the Gaussian uncertainty metric and the NVS selection module: both are algorithmic components, not entities with independent falsifiable handles, and both rest on the assumptions listed above. The main uncharged inputs are the three free operating points (view count, refinement iterations, metric weights).

free parameters (3)
  • Number of selected novel views = 100
    Table S5 sweeps 0/100/150 views; 100 is chosen as the best trade-off between rendering quality (improves with count) and reconstruction depth L1 plus FPS (both degrade at 150). Manually tuned on evaluation scenes, not a held-out set.
  • Final refinement iterations = 2,000
    Table S6 sweeps 2K/5K/10K/26K; 2K is chosen because geometric accuracy declines with more iterations and runtime suffers, while rendering keeps improving. A data-dependent operating point.
  • Uncertainty metric weighting (shape vs position-gradient terms) = not specified in provided text
    The uncertainty is defined from Gaussian shapes and positional gradients (contributions, bullet 2); the relative weights and normalization live in the method section, which is not present in the provided excerpt. Unknown whether they were tuned.
axioms (5)
  • domain assumption MVSFormer with DROID-SLAM disparity injection produces depths accurate enough to supervise and refine 3DGS
    The whole pipeline and the final refinement's depth loss depend on these estimated depths; accuracy is only indirectly validated via downstream metrics (e.g., Table S7).
  • domain assumption Online MVS (Song et al. 2021) provides consistent 3D information across frames and views
    Claimed in the Introduction and contribution bullets; the consistency guarantee is asserted, not derived, and the cited work is by the same authors.
  • domain assumption Color-only training on selected non-keyframes (no depth supervision) can improve scene completeness without degrading geometry
    NVS training uses frames without depth (Table S5 note); the paper itself observes reconstruction quality declines when more than 100 such views are added, so the assumption holds only within a narrow operating range.
  • domain assumption Global bundle adjustment plus Gaussian deformation maintains frontend-backend consistency
    Introduced in Section 1 as a design guarantee; no derivation or isolated ablation is present in the provided text.
  • standard math Standard 3DGS optimization (Kerbl et al. 2023) with color and depth losses converges to the stated quality under real-time constraints
    Base framework taken from prior literature; not re-derived here.

reviewed 2026-08-05 · how reviews work

0 comments
Cite this review

Pith. "Pith review of Online 3D Gaussian Splatting Modeling with Novel View Selection." pith.science (2026). https://pith.science/paper/EQ4D6XGH

@misc{pith2026250814014,
  author       = {Pith},
  title        = {Pith review of: Online 3D Gaussian Splatting Modeling with Novel View Selection},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/EQ4D6XGH}},
  note         = {Machine review of arXiv:2508.14014}
}
Share X Bluesky LinkedIn Reddit HN
read the original abstract

This study addresses the challenge of generating online 3D Gaussian Splatting (3DGS) models from RGB-only frames. Previous studies have employed dense SLAM techniques to estimate 3D scenes from keyframes for 3DGS model construction. However, these methods are limited by their reliance solely on keyframes, which are insufficient to capture an entire scene, resulting in incomplete reconstructions. Moreover, building a generalizable model requires incorporating frames from diverse viewpoints to achieve broader scene coverage. However, online processing restricts the use of many frames or extensive training iterations. Therefore, we propose a novel method for high-quality 3DGS modeling that improves model completeness through adaptive view selection. By analyzing reconstruction quality online, our approach selects optimal non-keyframes for additional training. By integrating both keyframes and selected non-keyframes, the method refines incomplete regions from diverse viewpoints, significantly enhancing completeness. We also present a framework that incorporates an online multi-view stereo approach, ensuring consistency in 3D information throughout the 3DGS modeling process. Experimental results demonstrate that our method outperforms state-of-the-art methods, delivering exceptional performance in complex outdoor scenes.

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

44 extracted references · 37 canonical work pages · 1 internal anchor

  1. [1]

    Zoedepth: Zero- shot transfer by combining relative and metric depth.arXiv preprint arXiv:2302.12288,

    [Bhatet al., 2023 ] Shariq Farooq Bhat, Reiner Birkl, Diana Wofk, Peter Wonka, and Matthias M¨uller. Zoedepth: Zero- shot transfer by combining relative and metric depth.arXiv preprint arXiv:2302.12288,

    Pith/arXiv arXiv 2023

  2. [7]

    Naruto: Neural active reconstruc- tion from uncertain target observations

    [Fenget al., 2024 ] Ziyue Feng, Huangying Zhan, Zheng Chen, Qingan Yan, Xiangyu Xu, Changjiang Cai, Bing Li, Qilun Zhu, and Yi Xu. Naruto: Neural active reconstruc- tion from uncertain target observations. InProceedings of the IEEE/CVF Conference on Computer Vision and Pat- tern Recognition, pages 21572–21583,

    work page 2024

  3. [10]

    Mgso: Monocular real-time photometric slam with efficient 3d gaussian splatting.arXiv preprint arXiv:2409.13055,

    [Huet al., 2024 ] Yan Song Hu, Nicolas Abboud, Muham- mad Qasim Ali, Adam Srebrnjak Yang, Imad Elhajj, Daniel Asmar, Yuhao Chen, and John S Zelek. Mgso: Monocular real-time photometric slam with efficient 3d gaussian splatting.arXiv preprint arXiv:2409.13055,

    work page arXiv 2024

  4. [11]

    Cg-slam: Efficient dense rgb-d slam in a consistent uncertainty-aware 3d gaussian field

    [Huet al., 2025 ] Jiarui Hu, Xianhao Chen, Boyin Feng, Guanglin Li, Liangjing Yang, Hujun Bao, Guofeng Zhang, and Zhaopeng Cui. Cg-slam: Efficient dense rgb-d slam in a consistent uncertainty-aware 3d gaussian field. InEu- ropean Conference on Computer Vision, pages 93–112. Springer,

    work page 2025

  5. [13]

    Fisherrf: Active view selection and mapping with radiance fields using fisher information

    [Jianget al., 2025 ] Wen Jiang, Boshu Lei, and Kostas Dani- ilidis. Fisherrf: Active view selection and mapping with radiance fields using fisher information. InEuropean Con- ference on Computer Vision, pages 422–440. Springer,

    work page 2025

  6. [14]

    Gs-planner: A gaussian-splatting-based planning framework for ac- tive high-fidelity reconstruction

    [Jinet al., 2024 ] Rui Jin, Yuman Gao, Yingjian Wang, Yuze Wu, Haojian Lu, Chao Xu, and Fei Gao. Gs-planner: A gaussian-splatting-based planning framework for ac- tive high-fidelity reconstruction. In2024 IEEE/RSJ In- ternational Conference on Intelligent Robots and Systems (IROS), pages 11202–11209. IEEE,

    work page 2024

  7. [15]

    3d gaus- sian splatting for real-time radiance field rendering.ACM Trans

    [Kerblet al., 2023 ] Bernhard Kerbl, Georgios Kopanas, Thomas Leimk ¨uhler, and George Drettakis. 3d gaus- sian splatting for real-time radiance field rendering.ACM Trans. Graph., 42(4):139–1,

    work page 2023

  8. [16]

    Tanks and temples: Bench- marking large-scale scene reconstruction.ACM Transac- tions on Graphics (ToG), 36(4):1–13,

    [Knapitschet al., 2017 ] Arno Knapitsch, Jaesik Park, Qian- Yi Zhou, and Vladlen Koltun. Tanks and temples: Bench- marking large-scale scene reconstruction.ACM Transac- tions on Graphics (ToG), 36(4):1–13,

    work page 2017

  9. [17]

    Tandem: Tracking and dense mapping in real-time using deep multi-view stereo

    [Koestleret al., 2022 ] Lukas Koestler, Nan Yang, Niclas Zeller, and Daniel Cremers. Tandem: Tracking and dense mapping in real-time using deep multi-view stereo. InConference on Robot Learning, pages 34–45. PMLR,

    work page 2022

  10. [18]

    Mvs-gs: High-quality 3d gaussian splatting mapping via online multi-view stereo.arXiv preprint arXiv:2412.19130,

    [Leeet al., 2024 ] Byeonggwon Lee, Junkyu Park, Khang Truong Giang, Sungho Jo, and Soohwan Song. Mvs-gs: High-quality 3d gaussian splatting mapping via online multi-view stereo.arXiv preprint arXiv:2412.19130,

    Pith/arXiv arXiv 2024

  11. [19]

    Fastscene: Text-driven fast 3d indoor scene generation via panoramic gaussian splatting

    [Maet al., 2024 ] Yikun Ma, Dandan Zhan, and Zhi Jin. Fastscene: Text-driven fast 3d indoor scene generation via panoramic gaussian splatting. InIJCAI,

    work page 2024

  12. [20]

    Gaussian splatting slam

    [Matsukiet al., 2024 ] Hidenobu Matsuki, Riku Murai, Paul HJ Kelly, and Andrew J Davison. Gaussian splatting slam. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18039–18048,

    work page 2024

  13. [21]

    Nerf: Representing scenes as neural radiance fields for view synthesis.Communications of the ACM, 65(1):99–106,

    [Mildenhallet al., 2021 ] Ben Mildenhall, Pratul P Srini- vasan, Matthew Tancik, Jonathan T Barron, Ravi Ra- mamoorthi, and Ren Ng. Nerf: Representing scenes as neural radiance fields for view synthesis.Communications of the ACM, 65(1):99–106,

    2021

  14. [22]

    Orb-slam: a versatile and accurate monocular slam system.IEEE transactions on robotics, 31(5):1147–1163,

    [Mur-Artalet al., 2015 ] Raul Mur-Artal, Jose Maria Mar- tinez Montiel, and Juan D Tardos. Orb-slam: a versatile and accurate monocular slam system.IEEE transactions on robotics, 31(5):1147–1163,

    work page 2015

  15. [25]

    Neurar: Neural uncertainty for autonomous 3d re- construction with implicit neural representations.IEEE Robotics and Automation Letters, 8(2):1125–1132,

    [Ranet al., 2023 ] Yunlong Ran, Jing Zeng, Shibo He, Jim- ing Chen, Lincheng Li, Yingfeng Chen, Gimhee Lee, and Qi Ye. Neurar: Neural uncertainty for autonomous 3d re- construction with implicit neural representations.IEEE Robotics and Automation Letters, 8(2):1125–1132,

    work page 2023

  16. [26]

    Vision transformers for dense prediction

    [Ranftlet al., 2021 ] Ren´e Ranftl, Alexey Bochkovskiy, and Vladlen Koltun. Vision transformers for dense prediction. InProceedings of the IEEE/CVF international conference on computer vision, pages 12179–12188,

    work page 2021

  17. [27]

    You only look once: Unified, real-time object detection

    [Redmon, 2016] J Redmon. You only look once: Unified, real-time object detection. InProceedings of the IEEE conference on computer vision and pattern recognition,

    work page 2016

  18. [29]

    Revising densification in gaussian splatting

    [Rota Bul`oet al., 2024 ] Samuel Rota Bul `o, Lorenzo Porzi, and Peter Kontschieder. Revising densification in gaussian splatting. InEuropean Conference on Computer Vision, pages 347–362. Springer,

    work page 2024

  19. [30]

    Splat-slam: Globally optimized rgb-only slam with 3d gaussians.arXiv preprint arXiv:2405.16544,

    [Sandstr¨omet al., 2024 ] Erik Sandstr ¨om, Keisuke Tateno, Michael Oechsle, Michael Niemeyer, Luc Van Gool, Mar- tin R Oswald, and Federico Tombari. Splat-slam: Globally optimized rgb-only slam with 3d gaussians.arXiv preprint arXiv:2405.16544,

    Pith/arXiv arXiv 2024

  20. [31]

    View planning for automated three- dimensional object reconstruction and inspection.ACM Computing Surveys (CSUR), 35(1):64–96,

    [Scottet al., 2003 ] William R Scott, Gerhard Roth, and Jean- Franc ¸ois Rivest. View planning for automated three- dimensional object reconstruction and inspection.ACM Computing Surveys (CSUR), 35(1):64–96,

    work page 2003

  21. [33]

    The replica dataset: A digital replica of indoor spaces.arXiv preprint arXiv:1906.05797,

    [Straubet al., 2019 ] Julian Straub, Thomas Whelan, Lingni Ma, Yufan Chen, Erik Wijmans, Simon Green, Jakob J En- gel, Raul Mur-Artal, Carl Ren, Shobhit Verma, et al. The replica dataset: A digital replica of indoor spaces.arXiv preprint arXiv:1906.05797,

    Pith/arXiv arXiv 2019

  22. [37]

    Bovik, H.R

    [Wanget al., 2004 ] Zhou Wang, A.C. Bovik, H.R. Sheikh, and E.P. Simoncelli. Image quality assessment: from error visibility to structural similarity.IEEE Transactions on Image Processing, 13(4):600–612,

    work page 2004

  23. [39]

    Active implicit ob- ject reconstruction using uncertainty-guided next-best- view optimization.IEEE Robotics and Automation Let- ters,

    [Yanet al., 2023 ] Dongyu Yan, Jianheng Liu, Fengyu Quan, Haoyao Chen, and Mengmeng Fu. Active implicit ob- ject reconstruction using uncertainty-guided next-best- view optimization.IEEE Robotics and Automation Let- ters,

    work page 2023

  24. [40]

    Mvsnet: Depth inference for unstructured multi-view stereo

    [Yaoet al., 2018 ] Yao Yao, Zixin Luo, Shiwei Li, Tian Fang, and Long Quan. Mvsnet: Depth inference for unstructured multi-view stereo. InProceedings of the European confer- ence on computer vision (ECCV), pages 767–783,

    work page 2018

  25. [41]

    Go-slam: Global optimiza- tion for consistent 3d instant reconstruction

    [Zhanget al., 2023 ] Youmin Zhang, Fabio Tosi, Stefano Mattoccia, and Matteo Poggi. Go-slam: Global optimiza- tion for consistent 3d instant reconstruction. InProceed- ings of the IEEE/CVF International Conference on Com- puter Vision, pages 3727–3737,

    work page 2023

  26. [42]

    Glorie-slam: Globally optimized rgb-only implicit encoding point cloud slam.arXiv preprint arXiv:2403.19549,

    [Zhanget al., 2024 ] Ganlin Zhang, Erik Sandstr¨om, Youmin Zhang, Manthan Patel, Luc Van Gool, and Martin R Oswald. Glorie-slam: Globally optimized rgb-only implicit encoding point cloud slam.arXiv preprint arXiv:2403.19549,

    Pith/arXiv arXiv 2024

  27. [43]

    RePaint-NeRF: Nerf editting via semantic masks and diffusion models

    [Zhouet al., 2023 ] Xingchen Zhou, Ying He, F Richard Yu, Jianqiang Li, and You Li. RePaint-NeRF: Nerf editting via semantic masks and diffusion models. InIJCAI,

    work page 2023

  28. [44]

    Nicer-slam: Neural implicit scene encoding for rgb slam

    [Zhuet al., 2024 ] Zihan Zhu, Songyou Peng, Viktor Lars- son, Zhaopeng Cui, Martin R Oswald, Andreas Geiger, and Marc Pollefeys. Nicer-slam: Neural implicit scene encoding for rgb slam. In2024 International Conference on 3D Vision (3DV), pages 42–52. IEEE, 2024

    work page 2024

  29. [2003]

    View path planning via online multiview stereo for 3-d modeling of large-scale structures.IEEE Transactions on Robotics, 38(1):372–390,

    [Songet al., 2021 ] Soohwan Song, Daekyum Kim, and Sunghee Choi. View path planning via online multiview stereo for 3-d modeling of large-scale structures.IEEE Transactions on Robotics, 38(1):372–390,

    work page 2021

  30. [2004]

    Tartanair: A dataset to push the limits of visual slam

    [Wanget al., 2020 ] Wenshan Wang, Delong Zhu, Xiangwei Wang, Yaoyu Hu, Yuheng Qiu, Chen Wang, Yafei Hu, Ashish Kapoor, and Sebastian Scherer. Tartanair: A dataset to push the limits of visual slam. In2020 IEEE/RSJ International Conference on Intelligent Robots and Sys- tems (IROS), pages 4909–4916. IEEE,

    work page 2020

  31. [2011]

    Q-SLAM: Quadric Representations for Monocular SLAM

    [Penget al., 2024 ] Chensheng Peng, Chenfeng Xu, Yue Wang, Mingyu Ding, Heng Yang, Masayoshi Tomizuka, Kurt Keutzer, Marco Pavone, and Wei Zhan. Q-slam: Quadric representations for monocular slam.arXiv preprint arXiv:2403.08125,

    work page internal anchor Pith review Pith/arXiv arXiv 2024

  32. [2012]

    Raft: Recurrent all-pairs field transforms for optical flow

    [Teed and Deng, 2020] Zachary Teed and Jia Deng. Raft: Recurrent all-pairs field transforms for optical flow. In Computer Vision–ECCV 2020: 16th European Confer- ence, Glasgow, UK, August 23–28, 2020, Proceedings, Part II 16, pages 402–419. Springer,

    work page 2020

  33. [2014]

    Ges: General- ized exponential splatting for efficient radiance field ren- dering

    [Hamdiet al., 2024 ] Abdullah Hamdi, Luke Melas-Kyriazi, Jinjie Mai, Guocheng Qian, Ruoshi Liu, Carl V ondrick, Bernard Ghanem, and Andrea Vedaldi. Ges: General- ized exponential splatting for efficient radiance field ren- dering. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 19812– 19822,

    work page 2024

  34. [2015]

    Dtam: Dense tracking and mapping in real-time

    [Newcombeet al., 2011 ] Richard A Newcombe, Steven J Lovegrove, and Andrew J Davison. Dtam: Dense tracking and mapping in real-time. In2011 international confer- ence on computer vision, pages 2320–2327. IEEE,

    work page 2011

  35. [2016]

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

    [Rosinolet al., 2023 ] Antoni Rosinol, John J Leonard, and Luca Carlone. Nerf-slam: Real-time dense monocular slam with neural radiance fields. In2023 IEEE/RSJ In- ternational Conference on Intelligent Robots and Systems (IROS), pages 3437–3444. IEEE,

    work page 2023

  36. [2017]

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

    [Eftekharet al., 2021 ] Ainaz Eftekhar, Alexander Sax, Jiten- dra Malik, and Amir Zamir. Omnidata: A scalable pipeline for making multi-task mid-level vision datasets from 3d scans. InProceedings of the IEEE/CVF International Con- ference on Computer Vision, pages 10786–10796,

    work page 2021

  37. [2018]

    Mvsformer: Multi-view stereo by learning robust image features and temperature-based depth.arXiv preprint arXiv:2208.02541,

    [Caoet al., 2022 ] Chenjie Cao, Xinlin Ren, and Yanwei Fu. Mvsformer: Multi-view stereo by learning robust image features and temperature-based depth.arXiv preprint arXiv:2208.02541,

    Pith/arXiv arXiv 2022

  38. [2019]

    A bench- mark for the evaluation of rgb-d slam systems

    [Sturmet al., 2012 ] J¨urgen Sturm, Nikolas Engelhard, Felix Endres, Wolfram Burgard, and Daniel Cremers. A bench- mark for the evaluation of rgb-d slam systems. In2012 IEEE/RSJ international conference on intelligent robots and systems, pages 573–580. IEEE,

    work page 2012

  39. [2020]

    Droid- slam: Deep visual slam for monocular, stereo, and rgb-d cameras.Advances in neural information processing sys- tems, 34:16558–16569,

    [Teed and Deng, 2021] Zachary Teed and Jia Deng. Droid- slam: Deep visual slam for monocular, stereo, and rgb-d cameras.Advances in neural information processing sys- tems, 34:16558–16569,

    work page 2021

  40. [2021]

    Direct sparse odometry.IEEE trans- actions on pattern analysis and machine intelligence, 40(3):611–625,

    [Engelet al., 2017 ] Jakob Engel, Vladlen Koltun, and Daniel Cremers. Direct sparse odometry.IEEE trans- actions on pattern analysis and machine intelligence, 40(3):611–625,

    work page 2017

  41. [2022]

    Scannet: Richly-annotated 3d reconstructions of indoor scenes

    [Daiet al., 2017 ] Angela Dai, Angel X Chang, Manolis Savva, Maciej Halber, Thomas Funkhouser, and Matthias Nießner. Scannet: Richly-annotated 3d reconstructions of indoor scenes. InProceedings of the IEEE conference on computer vision and pattern recognition, pages 5828– 5839,

    work page 2017

  42. [2023]

    Codeslam—learning a compact, optimisable repre- sentation for dense visual slam

    [Bloeschet al., 2018 ] Michael Bloesch, Jan Czarnowski, Ronald Clark, Stefan Leutenegger, and Andrew J Davi- son. Codeslam—learning a compact, optimisable repre- sentation for dense visual slam. InProceedings of the IEEE conference on computer vision and pattern recog- nition, pages 2560–2568,

    work page 2018

  43. [2024]

    Svo: Fast semi-direct monocular vi- sual odometry

    [Forsteret al., 2014 ] Christian Forster, Matia Pizzoli, and Davide Scaramuzza. Svo: Fast semi-direct monocular vi- sual odometry. In2014 IEEE international conference on robotics and automation (ICRA), pages 15–22. IEEE,

    work page 2014

  44. [2025]

    Photo-slam: Real-time si- multaneous localization and photorealistic mapping for monocular stereo and rgb-d cameras

    [Huanget al., 2024 ] Huajian Huang, Longwei Li, Hui Cheng, and Sai-Kit Yeung. Photo-slam: Real-time si- multaneous localization and photorealistic mapping for monocular stereo and rgb-d cameras. InProceedings of the IEEE/CVF Conference on Computer Vision and Pat- tern Recognition, pages 21584–21593,

    work page 2024

This paper was first reviewed by deepseek-v4-flash on August 5, 2026.