Pith. sign in

REVIEW 5 major objections 5 minor 69 references

RU4D-SLAM claims that reweighting per-pixel uncertainty, after integrating renders along the camera exposure path, makes 4D Gaussian Splatting SLAM accurate under motion, blur, and exposure inconsistencies.

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 ↗

T0 review

2026-08-02 21:12 UTC pith:XDBZDFG7

load-bearing objection Rendering results are strong, but the trajectory SOTA claim is a tie against a modified baseline and the uncertainty-mask logic is under-tested for blur-only inputs. the 5 major comments →

arxiv 2602.20807 v2 pith:XDBZDFG7 submitted 2026-02-24 cs.CV cs.RO

RU4D-SLAM: Reweighting Uncertainty in Gaussian Splatting SLAM for 4D Scene Reconstruction

Yangfan Zhao , Hanwei Zhang , Ke Huang , Qiufeng Wang , Zhenzhou Shao , Dengyu Wu This is my paper
classification cs.CV cs.RO
keywords 4D Gaussian splattingSLAMdynamic scenesuncertainty reweightingmotion blurexposure modelingscene reconstructionpose tracking
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.

This paper tries to establish that 4D Gaussian Splatting SLAM can be made robust to the two things that break current systems: objects moving through the scene, and low-quality images blurred or unevenly exposed by the camera's own motion. Its route is to treat per-pixel uncertainty not as a nuisance to be masked away but as a signal that, once reweighted by exposure modeling and semantic cues, locates exactly where the scene is deforming. On that basis it initializes local deformation anchors in the moving regions and learns time-varying opacities that let dynamic Gaussians fade in and out smoothly. If the claim holds, a single pipeline can track the camera and reconstruct a temporally coherent 4D scene directly from degraded monocular RGB input, without relying on object-class priors.

Core claim

The central claim is that uncertainty maps predicted for static reconstruction can be repurposed, with two corrective mechanisms, to drive 4D mapping: rendering is integrated along the camera trajectory during the exposure interval so that blur and exposure inconsistency no longer inflate per-pixel uncertainty; and the resulting uncertainty mask, merged with semantic segment candidates, isolates truly moving regions. Deformation nodes are seeded inside those regions and each node carries a learnable opacity trajectory, so that unreliable 2D pixel tracks cannot permanently warp Gaussians to wrong positions. The paper reports average PSNR of 25.95 dB on TUM, 26.33 dB on Bonn, and 24.22 dB on W

What carries the argument

The argument rests on three interlocking components. IR (integrate and render) approximates the camera exposure integral by summing renders at interpolated poses along the trajectory, with learned exposure gain and bias, so motion-blurred frames are compared against blurred renders rather than sharp ones; this stabilizes the per-pixel uncertainty map. RUM (reweighted uncertainty mask) thresholds that uncertainty map and merges any semantic segmentation candidate overlapping it, under the premise that high static-mapping uncertainty indicates motion, producing a binary mask that seeds deformation nodes. AOW (adaptive opacity weighting) attaches a learnable time-varying opacity to each deforma

Load-bearing premise

The mask that separates moving from static regions assumes that high uncertainty in the static map is caused by motion rather than by blur or exposure artifacts that IR failed to absorb; if a strongly blurred static background still yields high uncertainty, the deformation nodes will be seeded in the background and pull static Gaussians incorrectly.

What would settle it

A controlled sequence in which the camera moves while the scene is perfectly static, with blur artificially applied: if RUM seeds deformation nodes in the static background (i.e., the uncertainty mask exceeds the threshold in blurred regions), the core assumption fails. Quantitatively, compare ATE and rendered PSNR against the same method with the motion mask forced to empty; a meaningful drop in PSNR under pure blur would indicate the mask misfires.

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

Share X Bluesky LinkedIn Reddit HN

If this is right

  • Camera trajectories and 4D scene models can be estimated jointly from degraded monocular input, removing the need for object-class priors or explicit dynamic-object detection.
  • Blur and exposure inconsistency stop being failure modes: because rendered blur is matched to observed blur, static mapping can stay dense instead of masking out uncertain pixels.
  • Uncertainty maps become usable as motion detectors, so the same per-pixel reliability signal that stabilizes tracking can seed dynamic deformation.
  • Ablations indicate that the exposure-integration term alone nearly halves the number of Gaussians needed to explain a dynamic blurry scene.
  • If the reported numbers generalize, 4D Gaussian SLAM becomes practical for handheld and outdoor capture, not just clean RGB-D.

Where Pith is reading between the lines

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

  • The RUM design suggests a testable separation: if high uncertainty in a static, unblurred scene never triggers deformation, then thresholding on an exposure-aware uncertainty map could serve as a general motion proposal mechanism, replacing optical flow or object detectors.
  • Because the exposure integral is trained from the data itself, the same IR module might also absorb other photometric degradations such as defocus, rolling shutter, or sensor noise, as long as they can be approximated by integrating renders along a trajectory.
  • The learnable opacity per node could be interpreted as a continuous visibility field; a natural extension is to use it as a confidence for downstream tasks like segmentation, or as a cue to spawn and despawn Gaussians, reducing memory in long traversals.
  • If the hand-set thresholds (δ_u, δ_ru) were cross-validated or made adaptive, the method could extend to sequences with no moving objects, where currently the fixed threshold might produce spurious deformation seeds.

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

5 major / 5 minor

Summary. This paper presents RU4D-SLAM, a Gaussian-splatting SLAM system for dynamic scenes with low-quality inputs. The system builds on WildGS-SLAM's per-pixel uncertainty model and adds three components: an integrate-and-render (IR) module that accumulates rendering over an exposure interval to handle motion blur and exposure (Eqs. 4–6); a reweighted uncertainty mask (RUM) that thresholds the uncertainty map and merges SAM segments to identify dynamic regions (Eqs. 8–11); and adaptive opacity weighting (AOW) that learns time-varying visibility for deformation nodes (Eqs. 12–14). Deformation nodes are initialized inside RUM regions using SpaTracker and optimized with the MoSca graph. Experiments on TUM, Bonn, and Wild-SLAM report state-of-the-art rendering metrics and competitive ATE.

Significance. The paper addresses a real and timely problem: extending Gaussian-splatting SLAM to 4D reconstruction when the input contains moving objects, motion blur, and exposure variations. The proposed components are reasonable and the qualitative comparisons in Figs. 3–5 show visible improvements. The authors provide code, ablate IR and AOW quantitatively on TUM (Table 1), and evaluate on three public benchmarks. If the comparisons are made fair and the central RUM assumption is validated under blur-only conditions, the framework would be a useful contribution. However, the current evidence does not fully support the abstract's claim of substantially outperforming state-of-the-art methods in trajectory accuracy, and the core RUM segmentation is not stress-tested in the failure mode most relevant to low-quality inputs.

major comments (5)
  1. [Sec. 4.2, Eq. (11)] The sentence immediately after Eq. (11) states that 'high-uncertainty regions in static mapping typically indicate motion.' This assumption is load-bearing: M_ru seeds all deformation nodes (Sec. 4.3) and is formed once before 4D mapping. All evaluated sequences contain moving objects, so the blur-only failure mode is confounded. The IR model in Eqs. (4)–(6) integrates only camera motion over the exposure interval; it does not model object-motion blur or residual static blur when the pose is imperfect. If β² remains high for a static blurred region, Eq. (9) flags it, SAM can propagate the false positive, and spurious deformation nodes persist. Please provide a quantitative experiment on sequences with motion blur but no dynamic objects (e.g., TUM static sequences or synthetic blur applied to static frames) and report mask precision/recall and resulting ATE/PSNR. Also, ablate RUM by repla
  2. [Sec. 5.3, Tables 6–7] The trajectory-accuracy claim is overstated. Over WildGS-SLAM*, the average ATE improvement is 0.01 cm on TUM and 0.28 cm on Bonn; on TUM several per-sequence entries are within 0.01–0.05 cm of that baseline. Moreover, WildGS-SLAM* is a re-evaluated variant with the final bundle adjustment removed. Without multiple runs or error bars, this is a statistical tie, not 'substantially outperforms state-of-the-art approaches in trajectory accuracy' as claimed in the abstract. Please compare against the unmodified published WildGS-SLAM, report variance over multiple seeds, or temper the claim to 'competitive.'
  3. [Sec. 5.1, Tables 3–7] The depth protocol is not held constant across methods. The paper states 'Following WildGS-SLAM, we use RGB inputs and estimate depth using Metric3D across all datasets.' Several baselines (SplaTAM, Gaussian-SLAM, and MonoGS when used in RGB-D mode) are designed for sensor depth, and their published numbers may come from that setting. Training with Metric3D pseudo-depth as supervision (Eq. 7) can give the proposed method an advantage or disadvantage relative to published RGB-D numbers. Please re-run the baselines under the same monocular + Metric3D protocol, or clearly separate the input modality of each baseline and add a controlled comparison where all methods see the same depth source.
  4. [Sec. 5.2, Table 1] There is no quantitative ablation of RUM, despite RUM being one of the three named contributions. Table 1 removes only AOW and IR. The qualitative mask comparison in Figure 4 is suggestive but does not measure the effect of RUM on tracking or rendering. Add an ablation that removes RUM (e.g., seeding deformation nodes from the entire uncertainty mask, or from an optical-flow mask) and report PSNR, LPIPS, and ATE. Also report sensitivity to the hand-set thresholds δ_u=3.5 and δ_ru=0.2, since those hyperparameters are not derived from the problem.
  5. [Sec. 4.2, Eq. (8), and Sec. 4.3] There is a soft circularity in the RUM construction: the per-pixel uncertainty β² is trained with a photometric/depth objective (Eq. 8), RUM is thresholded from β², deformation nodes are seeded from RUM, and the deformation field is then optimized with the same photometric loss (Eq. 7). This makes the dynamic/static split partly a byproduct of the current photometric residual rather than an independent motion cue. The authors should either discuss this explicitly or provide an ablation with an independent motion signal (e.g., a frozen optical-flow network) to show that RUM adds information beyond reweighted photometric error.
minor comments (5)
  1. [Eq. (8)] The equation is ambiguous: it is unclear whether β_i² divides the entire expression (SSIM' + depth term) or only the depth term. Please add parentheses.
  2. [Sec. 4.3 and Fig. 2] The name of the tracking model is inconsistent: 'Spa-tracker' in the text versus 'SpaTracker' in Figure 2 and reference [58]. Please unify.
  3. [Sec. 5.1] The AOW regularization is described only verbally ('we add regularization to node ensure the similarity of AOW at similar times'). The exact regularizer should be given in equation form for reproducibility.
  4. [Sec. 4.2] Typos: 'hyperparamter' after Eq. (11); the sentence following Eq. (10) reads awkwardly. Minor language editing would help.
  5. [Abstract and Sec. 6] The abstract calls the framework 'efficient,' but no runtime or resource measurements are reported; the conclusion states that real-time performance remains future work. Please either report wall-clock/FPS numbers or soften the 'efficient' wording.

Circularity Check

0 steps flagged

No significant circularity; the uncertainty-to-motion link in RUM is a heuristic assumption, not an equationally forced reduction.

full rationale

The paper contains no load-bearing self-citations: the uncertainty formulation is attributed to WildGS-SLAM [66], the motion-scaffold graph to MoSca [23], and the exposure-integration idea to BAD-NeRF [52], none of which share authors with RU4D-SLAM. The apparent bootstrap in the RUM pipeline is not a circular reduction. Eq. (8) trains the per-pixel uncertainty β² as a robust loss weight against static rendering residuals; Eq. (9) thresholds it to form M_u; Eq. (11) merges this with SAM candidates; and Sec. 4.3 seeds deformation nodes inside RUM regions, whose dynamic Gaussians are then optimized with the photometric/depth loss in Eq. (7). This is a residual-driven segmentation followed by capacity allocation, a standard robust-estimation pattern, not a definitional identity: β² is not defined to be the residual by construction, it is a regularized learned quantity, and the final quality metrics (PSNR/SSIM/LPIPS, ATE) are evaluated against external data and compared with external baselines rather than against the mask itself. The stated premise that 'high-uncertainty regions in static mapping typically indicate motion' is a testable empirical assumption; its possible failure under blur-only inputs is a correctness risk, not evidence of circularity. IR and AOW are forward rendering/opacity mechanisms optimized on the same reconstruction losses, which is ordinary supervised fitting rather than a prediction derived from its own input. Therefore no step in the derivation chain reduces to its own outputs.

Axiom & Free-Parameter Ledger

5 free parameters · 6 axioms · 0 invented entities

The core novelty is algorithmic integration. The few newly introduced tunable parameters are mask thresholds and loss weights; the learned exposure and control poses are trained within the system. The assumptions above are inherited from the underlying pretrained models and deformation graph and are not independently validated here.

free parameters (5)
  • δ_u = 3.5
    Uncertainty mask threshold; manually chosen (Sec 5.1); determines what counts as dynamic region for RUM.
  • δ_ru = 0.2
    SAM mask merging threshold; manually chosen (Sec 5.1); controls which semantic candidates merge into RUM.
  • λ_1, λ_2 = not reported
    Loss weights in Eqs (7)-(8); balanced by hand, values not stated in the text.
  • a, b = learned (not reported)
    Exposure gain/bias in blur rendering (Eq 5); fitted on data during training.
  • T_s, T_e = initialized from pose + noise, optimized
    Control poses for blur interpolation (Eq 6); fitted during training.
axioms (6)
  • domain assumption High-uncertainty regions in static mapping typically indicate motion
    Sec 4.2; basis for constructing RUM; unverified for blur-only regions.
  • domain assumption Scene changes have minor effect during short exposure intervals
    Sec 4.1; justifies rendering static Gaussians to synthesize blurred frames.
  • domain assumption Metric3D monocular depth is accurate enough to supervise DBA and mapping in dynamic scenes
    Sec 5.1; depth uses Metric3D; errors in depth cause wrong DBA weights and mapping.
  • domain assumption MoSca deformation graph with dual-quaternion blending spans the scene motions
    Sec 3; borrowed from [23]; limits the expressiveness for large/rigid/fast motions.
  • domain assumption DINO-based MLP uncertainty is a valid per-pixel reliability measure
    Sec 3 and Eq 8; borrowed from WildGS; no analysis of calibration.
  • domain assumption Pretrained SpaTracker 2D tracks provide accurate node trajectory initialization
    Sec 4.3; acknowledged as fragile; AOW is designed to compensate, but tracking failures may still misinitialize nodes.

reviewed 2026-08-02 · how reviews work

0 comments
Cite this review

Pith. "Pith review of RU4D-SLAM: Reweighting Uncertainty in Gaussian Splatting SLAM for 4D Scene Reconstruction." pith.science (2026). https://pith.science/paper/XDBZDFG7

@misc{pith2026260220807,
  author       = {Pith},
  title        = {Pith review of: RU4D-SLAM: Reweighting Uncertainty in Gaussian Splatting SLAM for 4D Scene Reconstruction},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/XDBZDFG7}},
  note         = {Machine review of arXiv:2602.20807}
}
Share X Bluesky LinkedIn Reddit HN
read the original abstract

Combining 3D Gaussian splatting with Simultaneous Localization and Mapping (SLAM) has gained popularity as it enables continuous 3D environment reconstruction during motion. However, existing methods struggle in dynamic environments, particularly moving objects complicate 3D reconstruction and, in turn, hinder reliable tracking. The emergence of 4D reconstruction, especially 4D Gaussian splatting, offers a promising direction for addressing these challenges, yet its potential for 4D-aware SLAM remains largely underexplored. Along this direction, we propose a robust and efficient framework, namely Reweighting Uncertainty in Gaussian Splatting SLAM (RU4D-SLAM) for 4D scene reconstruction, that introduces temporal factors into spatial 3D representation while incorporating uncertainty-aware perception of scene changes, blurred image synthesis, and dynamic scene reconstruction. We enhance dynamic scene representation by integrating motion blur rendering, and improve uncertainty-aware tracking by extending per-pixel uncertainty modeling, which is originally designed for static scenarios, to handle blurred images. Furthermore, we propose a semantic-guided reweighting mechanism for per-pixel uncertainty estimation in dynamic scenes, and introduce a learnable opacity weight to support adaptive 4D mapping. Extensive experiments on standard benchmarks demonstrate that our method substantially outperforms state-of-the-art approaches in both trajectory accuracy and 4D scene reconstruction, particularly in dynamic environments with moving objects and low-quality inputs. Code available: https://ru4d-slam.github.io

Figures

Figures reproduced from arXiv: 2602.20807 by Dengyu Wu, Hanwei Zhang, Ke Huang, Qiufeng Wang, Yangfan Zhao, Zhenzhou Shao.

Figure 1
Figure 1. Figure 1: 4D scene reconstruction with RU4D-SLAM. The left side shows the 4D Gaussian map reconstructed by RU4D-SLAM on the Bonn dataset [45], featuring novel synthesized views that capture the temporal motions (top) and a comparison between the RGB input and RU4D-SLAM rendering at the same pose (bottom). On the right, we compare rendered results from MonoGS [32], 4DGS-SLAM [26], and RU4D-SLAM in dynamic scenes. The… view at source ↗
Figure 2
Figure 2. Figure 2: Overview of RU4D-SLAM. RU4D-SLAM operates in three stages: pose estimation, deformation field initialization, and 4D mapping, all of which are closely linked to the uncertainty map β 2 . In the pose estimation stage, β 2 supports uncertainty-aware DBA tracking. Before 4D mapping, the uncertainty map is combined with SAM to form RUM, within which deformation nodes are initialized as local motion anchors by … view at source ↗
Figure 3
Figure 3. Figure 3: Impact of IR on predicted uncertainty map β 2 . Pre￾dicted uncertainty on a sample frame from w x and w r sequence in TUM dataset. Higher values are shown in red, lower values in blue, with intermediate values transitioning smoothly. Qualitative Impact. To evaluate the impact of IR, we visualize the predicted uncertainty map β 2 from (8) with and without IR. As shown in [PITH_FULL_IMAGE:figures/full_fig_p… view at source ↗
Figure 4
Figure 4. Figure 4: Comparison of dynamic object tracking across masks. Performance of YOLO-based mask [26], flow-based mask [23], and our RUM on three sample frame sets from w x in TUM, pb1 in Bonn and park in Wild-SLAM. The performance of the RUM depends on the qual￾ity of the predicted uncertainty map, which benefits from the effectiveness of IR. To demonstrate RUM’s capability in tracking dynamic objects, we compare its p… view at source ↗
Figure 5
Figure 5. Figure 5: Impact of AOW on rendering. Comparison of render￾ings without and with AOW on frames from w x in TUM, pb1 in Bonn and park in Wild-SLAM. To evaluate the impact of AOW, we compare rendering results with and without AOW against the corresponding RGB inputs. As shown in the first column of [PITH_FULL_IMAGE:figures/full_fig_p007_5.png] view at source ↗

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

69 extracted references · 5 linked inside Pith

  1. [1]

    Dynaslam: Tracking, mapping, and inpainting in dynamic scenes.IEEE Robotics and Automation Letters, 3(4):4076– 4083, 2018

    Berta Bescos, Jos ´e M F ´acil, Javier Civera, and Jos ´e Neira. Dynaslam: Tracking, mapping, and inpainting in dynamic scenes.IEEE Robotics and Automation Letters, 3(4):4076– 4083, 2018. 2, 3

    2018

  2. [2]

    Dynaslam ii: Tightly-coupled multi-object tracking and slam.IEEE robotics and automation letters, 6(3):5191– 5198, 2021

    Berta Bescos, Carlos Campos, Juan D Tard ´os, and Jos ´e Neira. Dynaslam ii: Tightly-coupled multi-object tracking and slam.IEEE robotics and automation letters, 6(3):5191– 5198, 2021. 3

    2021

  3. [3]

    Hexplane: A fast representa- tion for dynamic scenes

    Ang Cao and Justin Johnson. Hexplane: A fast representa- tion for dynamic scenes. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 130–141, 2023. 3

    2023

  4. [4]

    Lsd- slam: Large-scale direct monocular slam

    Jakob Engel, Thomas Sch ¨ops, and Daniel Cremers. Lsd- slam: Large-scale direct monocular slam. InEuropean con- ference on computer vision, pages 834–849. Springer, 2014. 2

    2014

  5. [5]

    Direct sparse odometry.IEEE transactions on pattern analysis and machine intelligence, 40(3):611–625, 2017

    Jakob Engel, Vladlen Koltun, and Daniel Cremers. Direct sparse odometry.IEEE transactions on pattern analysis and machine intelligence, 40(3):611–625, 2017. 2

    2017

  6. [6]

    Fast dynamic radiance fields with time-aware neural vox- els

    Jiemin Fang, Taoran Yi, Xinggang Wang, Lingxi Xie, Xi- aopeng Zhang, Wenyu Liu, Matthias Nießner, and Qi Tian. Fast dynamic radiance fields with time-aware neural vox- els. InSIGGRAPH Asia 2022 Conference Papers, pages 1–9,

    2022

  7. [7]

    K-planes: Explicit radiance fields in space, time, and appearance

    Sara Fridovich-Keil, Giacomo Meanti, Frederik Rahbæk Warburg, Benjamin Recht, and Angjoo Kanazawa. K-planes: Explicit radiance fields in space, time, and appearance. In Proceedings of the IEEE/CVF Conference on Computer Vi- sion and Pattern Recognition, pages 12479–12488, 2023. 3

    2023

  8. [8]

    Learnable infinite taylor gaussian for dynamic view rendering

    Bingbing Hu, Yanyan Li, Rui Xie, Bo Xu, Haoye Dong, Junfeng Yao, and Gim Hee Lee. Learnable infinite taylor gaussian for dynamic view rendering. InProceedings of the Computer Vision and Pattern Recognition Conference, pages 26844–26854, 2025. 2, 3

    2025

  9. [9]

    Cg-slam: Efficient dense rgb-d slam in a consistent 1CPEC:https://perspicuous-computing.science uncertainty-aware 3d gaussian field

    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 1CPEC:https://perspicuous-computing.science uncertainty-aware 3d gaussian field. InEuropean Confer- ence on Computer Vision, pages 93–112. Springer, 2024. 3

    2024

  10. [10]

    Mu Hu, Wei Yin, Chi Zhang, Zhipeng Cai, Xiaoxiao Long, Hao Chen, Kaixuan Wang, Gang Yu, Chunhua Shen, and Shaojie Shen. Metric3d v2: A versatile monocular geomet- ric foundation model for zero-shot metric depth and surface normal estimation.IEEE Transactions on Pattern Analysis and Machine Intelligence, 2024. 5, 6

    2024

  11. [11]

    Photo-slam: Real-time simultaneous localization and photo- realistic mapping for monocular stereo and rgb-d cameras

    Huajian Huang, Longwei Li, Hui Cheng, and Sai-Kit Yeung. Photo-slam: Real-time simultaneous localization and photo- realistic mapping for monocular stereo and rgb-d cameras. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 21584–21593, 2024. 3

    2024

  12. [12]

    Sc-gs: Sparse-controlled gaussian splatting for editable dynamic scenes

    Yi-Hua Huang, Yang-Tian Sun, Ziyi Yang, Xiaoyang Lyu, Yan-Pei Cao, and Xiaojuan Qi. Sc-gs: Sparse-controlled gaussian splatting for editable dynamic scenes. InProceed- ings of the IEEE/CVF conference on computer vision and pattern recognition, pages 4220–4230, 2024. 2, 3, 8

    2024

  13. [13]

    Rodyn-slam: Robust dynamic dense rgb-d slam with neural radiance fields.IEEE Robotics and Automation Letters, 2024

    Haochen Jiang, Yueming Xu, Kejie Li, Jianfeng Feng, and Li Zhang. Rodyn-slam: Robust dynamic dense rgb-d slam with neural radiance fields.IEEE Robotics and Automation Letters, 2024. 3, 8

    2024

  14. [14]

    Neuman: Neural human radiance field from a single video

    Wei Jiang, Kwang Moo Yi, Golnoosh Samei, Oncel Tuzel, and Anurag Ranjan. Neuman: Neural human radiance field from a single video. InEuropean Conference on Computer Vision, pages 402–418. Springer, 2022. 2, 3

    2022

  15. [15]

    Eslam: Efficient dense slam system based on hybrid representation of signed distance fields

    Mohammad Mahdi Johari, Camilla Carta, and Franc ¸ois Fleuret. Eslam: Efficient dense slam system based on hybrid representation of signed distance fields. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 17408–17419, 2023. 2

    2023

  16. [16]

    Splatam: Splat track & map 3d gaussians for dense rgb-d slam

    Nikhil Keetha, Jay Karhade, Krishna Murthy Jatavallabhula, Gengshan Yang, Sebastian Scherer, Deva Ramanan, and Jonathon Luiten. Splatam: Splat track & map 3d gaussians for dense rgb-d slam. InProceedings of the IEEE/CVF Con- ference on Computer Vision and Pattern Recognition, pages 21357–21366, 2024. 3, 8

    2024

  17. [17]

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

    Bernhard Kerbl, Georgios Kopanas, Thomas Leimkuehler, and George Drettakis. 3d gaussian splatting for real-time radiance field rendering.ACM Trans. Graph., 42(4), 2023. 1, 3, 4

    2023

  18. [18]

    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. InProceedings of the IEEE/CVF International Con- ference on Computer Vision, pages 4015–4026, 2023. 5

    2023

  19. [19]

    Parallel tracking and map- ping for small ar workspaces

    Georg Klein and David Murray. Parallel tracking and map- ping for small ar workspaces. InISMAR, pages 225–234. IEEE, 2007. 2

    2007

  20. [20]

    Modec-gs: Global-to-local motion decomposition and temporal interval adjustment for compact dynamic 3d gaussian splatting

    Sangwoon Kwak, Joonsoo Kim, Jun Young Jeong, Won- Sik Cheong, Jihyong Oh, and Munchurl Kim. Modec-gs: Global-to-local motion decomposition and temporal interval adjustment for compact dynamic 3d gaussian splatting. In Proceedings of the Computer Vision and Pattern Recognition Conference, pages 11338–11348, 2025. 3

    2025

  21. [21]

    Efficient dynamic scene editing via 4d gaussian-based static-dynamic 9 separation

    Joohyun Kwon, Hanbyel Cho, and Junmo Kim. Efficient dynamic scene editing via 4d gaussian-based static-dynamic 9 separation. InProceedings of the Computer Vision and Pat- tern Recognition Conference, pages 26855–26865, 2025. 3

    2025

  22. [22]

    Fully explicit dynamic gaussian splat- ting.Advances in Neural Information Processing Systems, 37:5384–5409, 2024

    Junoh Lee, ChangYeon Won, Hyunjun Jung, Inhwan Bae, and Hae-Gon Jeon. Fully explicit dynamic gaussian splat- ting.Advances in Neural Information Processing Systems, 37:5384–5409, 2024. 3

    2024

  23. [23]

    Mosca: Dynamic gaussian fusion from casual videos via 4d motion scaffolds

    Jiahui Lei, Yijia Weng, Adam W Harley, Leonidas Guibas, and Kostas Daniilidis. Mosca: Dynamic gaussian fusion from casual videos via 4d motion scaffolds. InProceedings of the Computer Vision and Pattern Recognition Conference, pages 6165–6177, 2025. 2, 3, 6, 7

    2025

  24. [24]

    Ddn-slam: Real time dense dynamic neural implicit slam.IEEE Robotics and Au- tomation Letters, 2025

    Mingrui Li, Zhetao Guo, Tianchen Deng, Yiming Zhou, Yuxiang Ren, and Hongyu Wang. Ddn-slam: Real time dense dynamic neural implicit slam.IEEE Robotics and Au- tomation Letters, 2025. 3

    2025

  25. [25]

    Dy3dgs-slam: Monocular 3d gaussian splatting slam for dynamic environments.arXiv preprint arXiv:2506.05965,

    Mingrui Li, Yiming Zhou, Hongxing Zhou, Xinggang Hu, Florian Roemer, Hongyu Wang, and Ahmad Osman. Dy3dgs-slam: Monocular 3d gaussian splatting slam for dynamic environments.arXiv preprint arXiv:2506.05965,

    Pith/arXiv arXiv

  26. [26]

    4d gaussian splatting slam.arXiv preprint arXiv:2503.16710, 2025

    Yanyan Li, Youxu Fang, Zunjie Zhu, Kunyi Li, Yong Ding, and Federico Tombari. 4d gaussian splatting slam.arXiv preprint arXiv:2503.16710, 2025. 1, 2, 3, 7, 8

    Pith/arXiv arXiv 2025

  27. [27]

    Spacetime gaus- sian feature splatting for real-time dynamic view synthesis

    Zhan Li, Zhang Chen, Zhong Li, and Yi Xu. Spacetime gaus- sian feature splatting for real-time dynamic view synthesis. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8508–8520, 2024. 2, 3

    2024

  28. [28]

    Rgb-d inertial odometry for a resource-restricted robot in dy- namic environments.IEEE Robotics and Automation Letters, 7(4):9573–9580, 2022

    Jianheng Liu, Xuanfu Li, Yueqian Liu, and Haoyao Chen. Rgb-d inertial odometry for a resource-restricted robot in dy- namic environments.IEEE Robotics and Automation Letters, 7(4):9573–9580, 2022. 3

    2022

  29. [29]

    Robust dynamic radiance fields

    Yu-Lun Liu, Chen Gao, Andreas Meuleman, Hung-Yu Tseng, Ayush Saraf, Changil Kim, Yung-Yu Chuang, Jo- hannes Kopf, and Jia-Bin Huang. Robust dynamic radiance fields. InProceedings of the IEEE/CVF Conference on Com- puter Vision and Pattern Recognition, pages 13–23, 2023. 3

    2023

  30. [30]

    A survey: Learning embodied intelligence from physical simulators and world models.arXiv preprint arXiv:2507.00917, 2025

    Xiaoxiao Long, Qingrui Zhao, Kaiwen Zhang, Zihao Zhang, Dingrui Wang, Yumeng Liu, Zhengjie Shu, Yi Lu, Shouzheng Wang, Xinzhe Wei, et al. A survey: Learning embodied intelligence from physical simulators and world models.arXiv preprint arXiv:2507.00917, 2025. 2

    Pith/arXiv arXiv 2025

  31. [31]

    Dynamic 3d gaussians: Tracking by per- sistent dynamic view synthesis

    Jonathon Luiten, Georgios Kopanas, Bastian Leibe, and Deva Ramanan. Dynamic 3d gaussians: Tracking by per- sistent dynamic view synthesis. In2024 International Con- ference on 3D Vision (3DV), pages 800–809. IEEE, 2024. 3

    2024

  32. [32]

    Gaussian splatting slam

    Hidenobu Matsuki, Riku Murai, Paul HJ Kelly, and An- drew J Davison. Gaussian splatting slam. InProceedings of the IEEE/CVF Conference on Computer Vision and Pat- tern Recognition, pages 18039–18048, 2024. 1, 2, 3, 5, 8

    2024

  33. [33]

    Orb-slam: a versatile and accurate monocular slam system.IEEE Trans- actions on Robotics, 31(5):1147–1163, 2015

    Raul Mur-Artal and Juan D Tard ´os Montiel. Orb-slam: a versatile and accurate monocular slam system.IEEE Trans- actions on Robotics, 31(5):1147–1163, 2015. 1, 2

    2015

  34. [34]

    Kinectfusion: Real-time dense surface mapping and track- ing

    Richard A Newcombe, Shahram Izadi, Otmar Hilliges, David Molyneaux, David Kim, Andrew J Davison, Pushmeet Kohi, Jamie Shotton, Steve Hodges, and Andrew Fitzgibbon. Kinectfusion: Real-time dense surface mapping and track- ing. In2011 10th IEEE international symposium on mixed and augmented reality, pages 127–136. Ieee, 2011. 2

    2011

  35. [35]

    Dtam: Dense tracking and mapping in real-time

    Richard A Newcombe, Steven J Lovegrove, and Andrew J Davison. Dtam: Dense tracking and mapping in real-time. In2011 international conference on computer vision, pages 2320–2327. IEEE, 2011. 2

    2011

  36. [36]

    Refusion: 3d reconstruction in dynamic environments for rgb-d cameras exploiting resid- uals

    Emanuele Palazzolo, Jens Behley, Philipp Lottes, Philippe Gigu`ere, and Cyrill Stachniss. Refusion: 3d reconstruction in dynamic environments for rgb-d cameras exploiting resid- uals. In2019 IEEE/RSJ International Conference on Intel- ligent Robots and Systems (IROS), pages 546–553. IEEE,

  37. [37]

    Splinegs: Robust motion-adaptive spline for real-time dy- namic 3d gaussians from monocular video

    Jongmin Park, Minh-Quan Viet Bui, Juan Luis Gonza- lez Bello, Jaeho Moon, Jihyong Oh, and Munchurl Kim. Splinegs: Robust motion-adaptive spline for real-time dy- namic 3d gaussians from monocular video. InProceedings of the Computer Vision and Pattern Recognition Conference, pages 26866–26875, 2025. 2, 3

    2025

  38. [38]

    Desire-gs: 4d street gaussians for static-dynamic decomposition and surface reconstruction for urban driving scenes

    Chensheng Peng, Chengwei Zhang, Yixiao Wang, Chenfeng Xu, Yichen Xie, Wenzhao Zheng, Kurt Keutzer, Masayoshi Tomizuka, and Wei Zhan. Desire-gs: 4d street gaussians for static-dynamic decomposition and surface reconstruction for urban driving scenes. InProceedings of the Computer Vi- sion and Pattern Recognition Conference, pages 6782–6791,

  39. [39]

    D-nerf: Neural radiance fields for dynamic scenes

    Albert Pumarola, Enric Corona, Gerard Pons-Moll, and Francesc Moreno-Noguer. D-nerf: Neural radiance fields for dynamic scenes. InProceedings of the IEEE/CVF con- ference on computer vision and pattern recognition, pages 10318–10327, 2021. 2

    2021

  40. [40]

    Vins-mono: A robust and versatile monocular visual-inertial state estimator

    Tong Qin, Peiliang Li, and Shaojie Shen. Vins-mono: A robust and versatile monocular visual-inertial state estimator. IEEE transactions on robotics, 34(4):1004–1020, 2018. 2

    2018

  41. [41]

    Nerf on-the-go: Exploiting uncertainty for distractor-free nerfs in the wild

    Weining Ren, Zihan Zhu, Boyang Sun, Jiaqi Chen, Marc Pollefeys, and Songyou Peng. Nerf on-the-go: Exploiting uncertainty for distractor-free nerfs in the wild. InProceed- ings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8931–8940, 2024. 3, 5

    2024

  42. [42]

    Point-slam: Dense neural point cloud-based slam

    Erik Sandstr ¨om, Yue Li, Luc Van Gool, and Martin R Os- wald. Point-slam: Dense neural point cloud-based slam. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 18433–18444, 2023. 2

    2023

  43. [43]

    Lvi-sam: Tightly-coupled lidar-visual-inertial odometry via smoothing and mapping

    Tixiao Shan, Brendan Englot, Carlo Ratti, and Daniela Rus. Lvi-sam: Tightly-coupled lidar-visual-inertial odometry via smoothing and mapping. In2021 IEEE international confer- ence on robotics and automation (ICRA), pages 5692–5698. IEEE, 2021. 2

    2021

  44. [44]

    Nerf- player: A streamable dynamic scene representation with de- composed neural radiance fields.IEEE Transactions on Visu- alization and Computer Graphics, 29(5):2732–2742, 2023

    Liangchen Song, Anpei Chen, Zhong Li, Zhang Chen, Lele Chen, Junsong Yuan, Yi Xu, and Andreas Geiger. Nerf- player: A streamable dynamic scene representation with de- composed neural radiance fields.IEEE Transactions on Visu- alization and Computer Graphics, 29(5):2732–2742, 2023. 3

    2023

  45. [45]

    Bonn rgb-d dynamic dataset.https://www.ipb.uni-bonn.de/data/ rgbd-dynamic-dataset/, 2019

    StachnissLab, University of Bonn. Bonn rgb-d dynamic dataset.https://www.ipb.uni-bonn.de/data/ rgbd-dynamic-dataset/, 2019. Accessed: 2025-10-

    2019

  46. [46]

    Real-time visual odometry from dense rgb-d images

    Frank Steinbr ¨ucker, J ¨urgen Sturm, and Daniel Cremers. Real-time visual odometry from dense rgb-d images. In2011 IEEE international conference on computer vision work- shops (ICCV Workshops), pages 719–722. IEEE, 2011. 2

    2011

  47. [47]

    Sturm, N

    J. Sturm, N. Engelhard, F. Endres, W. Burgard, and D. Cre- mers. A benchmark for the evaluation of rgb-d slam systems. InProc. of the International Conference on Intelligent Robot Systems (IROS), 2012. 6

    2012

  48. [48]

    imap: Implicit mapping and positioning in real-time

    Edgar Sucar, Shikun Liu, Joseph Ortiz, and Andrew J Davi- son. imap: Implicit mapping and positioning in real-time. In Proceedings of the IEEE/CVF international conference on computer vision, pages 6229–6238, 2021. 2

    2021

  49. [49]

    DROID-SLAM: Deep visual SLAM for monocular, stereo, and RGB-d cameras

    Zachary Teed and Jia Deng. DROID-SLAM: Deep visual SLAM for monocular, stereo, and RGB-d cameras. InAd- vances in Neural Information Processing Systems, 2021. 2, 4

    2021

  50. [50]

    Non- rigid neural radiance fields: Reconstruction and novel view synthesis of a dynamic scene from monocular video

    Edgar Tretschk, Ayush Tewari, Vladislav Golyanik, Michael Zollh¨ofer, Christoph Lassner, and Christian Theobalt. Non- rigid neural radiance fields: Reconstruction and novel view synthesis of a dynamic scene from monocular video. InPro- ceedings of the IEEE/CVF international conference on com- puter vision, pages 12959–12970, 2021. 2

    2021

  51. [51]

    Co- slam: Joint coordinate and sparse parametric encodings for neural real-time slam

    Hengyi Wang, Jingwen Wang, and Lourdes Agapito. Co- slam: Joint coordinate and sparse parametric encodings for neural real-time slam. InProceedings of the IEEE/CVF Con- ference on Computer Vision and Pattern Recognition, pages 13293–13302, 2023. 2

    2023

  52. [52]

    BAD-NeRF: Bundle Adjusted Deblur Neural Radiance Fields

    Peng Wang, Lingzhe Zhao, Ruijie Ma, and Peidong Liu. BAD-NeRF: Bundle Adjusted Deblur Neural Radiance Fields. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 4170–4179, 2023. 1, 5

    2023

  53. [53]

    Dust3r: Geometric 3d vi- sion made easy

    Shuzhe Wang, Vincent Leroy, Yohann Cabon, Boris Chidlovskii, and Jerome Revaud. Dust3r: Geometric 3d vi- sion made easy. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 20697– 20709, 2024. 3

    2024

  54. [54]

    Freetimegs: Free gaussian primitives at anytime any- where for dynamic scene reconstruction

    Yifan Wang, Peishan Yang, Zhen Xu, Jiaming Sun, Zhan- hua Zhang, Yong Chen, Hujun Bao, Sida Peng, and Xiaowei Zhou. Freetimegs: Free gaussian primitives at anytime any- where for dynamic scene reconstruction. InProceedings of the Computer Vision and Pattern Recognition Conference, pages 21750–21760, 2025. 2

    2025

  55. [55]

    Hu- mannerf: Free-viewpoint rendering of moving people from monocular video

    Chung-Yi Weng, Brian Curless, Pratul P Srinivasan, Jonathan T Barron, and Ira Kemelmacher-Shlizerman. Hu- mannerf: Free-viewpoint rendering of moving people from monocular video. InProceedings of the IEEE/CVF con- ference on computer vision and pattern Recognition, pages 16210–16220, 2022. 2, 3

    2022

  56. [56]

    4d gaussian splatting for real-time dynamic scene rendering

    Guanjun Wu, Taoran Yi, Jiemin Fang, Lingxi Xie, Xiaopeng Zhang, Wei Wei, Wenyu Liu, Qi Tian, and Xinggang Wang. 4d gaussian splatting for real-time dynamic scene rendering. InProceedings of the IEEE/CVF conference on computer vi- sion and pattern recognition, pages 20310–20320, 2024. 2, 3

    2024

  57. [57]

    Add-slam: Adaptive dy- namic dense slam with gaussian splatting.arXiv preprint arXiv:2505.19420, 2025

    Wenhua Wu, Chenpeng Su, Siting Zhu, Tianchen Deng, Zhe Liu, and Hesheng Wang. Add-slam: Adaptive dy- namic dense slam with gaussian splatting.arXiv preprint arXiv:2505.19420, 2025. 3

    arXiv 2025

  58. [58]

    Spatialtracker: Tracking any 2d pixels in 3d space

    Yuxi Xiao, Qianqian Wang, Shangzhan Zhang, Nan Xue, Sida Peng, Yujun Shen, and Xiaowei Zhou. Spatialtracker: Tracking any 2d pixels in 3d space. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 20406–20417, 2024. 5

    2024

  59. [59]

    DG-SLAM: Robust dynamic gaus- sian splatting SLAM with hybrid pose optimization

    Yueming Xu, Haochen Jiang, Zhongyang Xiao, Jianfeng Feng, and Li Zhang. DG-SLAM: Robust dynamic gaus- sian splatting SLAM with hybrid pose optimization. In The Thirty-eighth Annual Conference on Neural Information Processing Systems, 2024. 2, 3, 8

    2024

  60. [60]

    Gs-slam: Dense visual slam with 3d gaussian splatting

    Chi Yan, Delin Qu, Dan Xu, Bin Zhao, Zhigang Wang, Dong Wang, and Xuelong Li. Gs-slam: Dense visual slam with 3d gaussian splatting. InProceedings of the IEEE/CVF Con- ference on Computer Vision and Pattern Recognition, pages 19595–19604, 2024. 2, 3

    2024

  61. [61]

    Banmo: Building animatable 3d neural models from many casual videos

    Gengshan Yang, Minh V o, Natalia Neverova, Deva Ra- manan, Andrea Vedaldi, and Hanbyul Joo. Banmo: Building animatable 3d neural models from many casual videos. In Proceedings of the IEEE/CVF Conference on Computer Vi- sion and Pattern Recognition, pages 2863–2873, 2022. 2

    2022

  62. [62]

    V ox-fusion: Dense tracking and mapping with voxel-based neural implicit representation

    Xingrui Yang, Hai Li, Hongjia Zhai, Yuhang Ming, Yuqian Liu, and Guofeng Zhang. V ox-fusion: Dense tracking and mapping with voxel-based neural implicit representation. In 2022 IEEE International Symposium on Mixed and Aug- mented Reality (ISMAR), pages 499–507. IEEE, 2022. 2

    2022

  63. [63]

    Deformable 3d gaussians for high- fidelity monocular dynamic scene reconstruction

    Ziyi Yang, Xinyu Gao, Wen Zhou, Shaohui Jiao, Yuqing Zhang, and Xiaogang Jin. Deformable 3d gaussians for high- fidelity monocular dynamic scene reconstruction. InPro- ceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 20331–20341, 2024. 3

    2024

  64. [64]

    Gaussian-slam: Photo-realistic dense slam with gaus- sian splatting.arXiv preprint arXiv:2312.10070, 2023

    Vladimir Yugay, Yue Li, Theo Gevers, and Martin R Os- wald. Gaussian-slam: Photo-realistic dense slam with gaus- sian splatting.arXiv preprint arXiv:2312.10070, 2023. 8

    Pith/arXiv arXiv 2023

  65. [65]

    Monst3r: A simple approach for estimat- ing geometry in the presence of motion.arXiv preprint arxiv:2410.03825, 2024

    Junyi Zhang, Charles Herrmann, Junhwa Hur, Varun Jam- pani, Trevor Darrell, Forrester Cole, Deqing Sun, and Ming- Hsuan Yang. Monst3r: A simple approach for estimat- ing geometry in the presence of motion.arXiv preprint arxiv:2410.03825, 2024. 3

    Pith/arXiv arXiv 2024

  66. [66]

    Wildgs-slam: Monocular gaussian splatting slam in dynamic environments

    Jianhao Zheng, Zihan Zhu, Valentin Bieri, Marc Pollefeys, Songyou Peng, and Iro Armeni. Wildgs-slam: Monocular gaussian splatting slam in dynamic environments. InPro- ceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 11461–11471, 2025. 2, 3, 4, 5, 6, 8

    2025

  67. [67]

    Up-slam: Adaptively structured gaussian slam with uncertainty prediction in dynamic environments.arXiv preprint arXiv:2505.22335, 2025

    Wancai Zheng, Linlin Ou, Jiajie He, Libo Zhou, Xinyi Yu, and Yan Wei. Up-slam: Adaptively structured gaussian slam with uncertainty prediction in dynamic environments.arXiv preprint arXiv:2505.22335, 2025. 3

    arXiv 2025

  68. [68]

    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. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 21634–21643, 2024. 1

    2024

  69. [69]

    Oswald, and Marc Polle- 11 feys

    Zihan Zhu, Songyou Peng, Viktor Larsson, Weiwei Xu, Hu- jun Bao, Zhaopeng Cui, Martin R. Oswald, and Marc Polle- 11 feys. Nice-slam: Neural implicit scalable encoding for slam. InProceedings of the IEEE/CVF Conference on Com- puter Vision and Pattern Recognition (CVPR), pages 12786– 12796, 2022. 2 12

    2022

This paper was first reviewed by deepseek-v4-flash on August 2, 2026.