Pith. sign in

REVIEW 5 major objections 5 minor 2 cited by

Dy3DGS-SLAM: Monocular 3D Gaussian Splatting SLAM for Dynamic Environments

T0 review · 5 major / 5 minor · reviewed 2026-08-07 · deepseek-v4-flash

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

Pith's one-line read This paper claims the first RGB-only 3D Gaussian Splatting SLAM system for dynamic scenes, matching or beating RGB-D baselines on the TUM and BONN datasets.

desk verdict A genuinely novel monocular dynamic 3DGS-SLAM system, but the undefined scale factor S_n leaves the central metric-scale claim unsubstantiated. read the letter →

arxiv 2506.05965 v1 pith:XA7XJ75X submitted 2025-06-06 cs.CV

Mingrui Li , Yiming Zhou , Hongxing Zhou , Xinggang Hu , Florian Roemer , Hongyu Wang , Ahmad Osman This is my paper
classification cs.CV
keywords dynamicSLAM3DGaussiansplattingmonocularRGBopticalflowdepthestimationmaskfusionposerenderingloss
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

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

The reading

Dy3DGS-SLAM sets out to prove that a single moving camera, with no depth sensor, can perform simultaneous localization and mapping in scenes full of moving objects. The paper claims to be the first 3D Gaussian Splatting SLAM system for dynamic environments using only monocular RGB input. It fuses optical-flow motion masks with monocular-depth masks through a Bayesian model, then uses the fused mask to constrain pose estimation and to penalize Gaussians belonging to dynamic pixels. On the TUM and BONN dynamic datasets the system reports tracking errors and rendering quality that match or beat RGB-D-based NeRF and 3DGS SLAM methods. If correct, this would let dynamic-scene SLAM run on ordinary cameras rather than depth sensors.

What carries the argument

The load-bearing mechanism is the fused dynamic mask $\hat{M}$. Optical-flow motion pixels are grouped by K-means into $k$ object clusters, and for each pixel a Bayesian posterior combines the depth mask $D_m$ and the flow mask $F_m$, binarizing at threshold $T=0.95$. That mask is then used twice: in tracking it zeroes out dynamic regions of the optical flow and injects a per-frame scale factor $S_n$ into the motion loss, recovering metric scale from monocular depth; in mapping it reweights the color and depth rendering losses and prunes Gaussians labeled dynamic by setting their depth to infinity. The scene itself is represented by 3D Gaussian ellipsoids rendered through splatting, with a convolutional pose network following the design of [27]. All reported gains trace back to the fused mask: it is what lets a single network iteration replace the three-iteration refinement of the baseline it improves on.

What would settle it

Feed the system a monocular sequence with a known ground-truth trajectory, then multiply every estimated depth map by 0.8 before it enters the pipeline; if the per-frame scale factor in the motion loss does not restore the original trajectory, the claimed scale-recovery mechanism is not doing the work.

Watch

Extended reading notes

Core claim

The central discovery is that dynamic-object interference in monocular 3DGS-SLAM can be removed in a single network pass by treating optical flow and monocular depth as conditionally independent evidence about which pixels are moving. A K-means clustering step separates moving objects, a Bayesian posterior fusion produces a binary mask, and a per-frame scale factor $S_n$ is inserted into the motion loss to recover metric scale from scale-ambiguous monocular depth. The same mask drives rendering losses that set dynamic Gaussians to infinite depth and reweight photometric and depth terms, eliminating floaters and ghosting. The paper reports an average ATE of 4.5 cm on the BONN dataset and 4.7 cm on TUM, and argues this matches or outperforms the compared RGB-D methods on tracking, with rendering comparisons shown qualitatively.

Load-bearing premise

The system assumes that a hidden per-frame scale factor makes monocular depth estimates line up with true real-world scale, but the paper never says how that factor is found; if it is wrong, both tracking scale and reconstructed geometry break.

Editorial extensions

If this is right

  • A depth sensor is no longer required for dynamic-scene dense SLAM: a monocular camera can produce metric-scale trajectories and clean reconstructions on the tested indoor datasets.
  • The single-iteration mask fusion makes the system faster than multi-iteration dynamic visual-odometry baselines, running at 17 FPS tracking with 10.3 ms network updates in the reported experiments.
  • Dynamic Gaussians are pruned by setting their depth to infinity, which removes floaters from reconstructed meshes and improves multi-view consistency of the static background.
  • Tracking on the BONN dataset averages 4.5 cm ATE, lower than the reported average of every RGB-D baseline compared in that evaluation.
  • Rendering quality on crowd and person-tracking sequences is higher than RGB-D NeRF and 3DGS baselines, with fewer transient artifacts in the visualized meshes.

Reading between the lines

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

  • The paper never specifies how the per-frame scale factor $S_n$ is obtained; a cautious extension would test whether it is computed from the monocular depth prior, from fixed calibration, or by optimization, since the motion loss depends on it.
  • Because the mask fusion assumes moving objects can be separated into spatially distinct K-means clusters, scenes with interleaved or heavily overlapping dynamic objects may break the posterior fusion; that is a testable stress case beyond the paper's indoor experiments.
  • The method's scale recovery is validated only on indoor datasets with ground-truth trajectories; applying the same system outdoors, where monocular depth estimators are less reliable, would likely require retraining or explicit scale supervision.
  • A natural next experiment is to replace the monocular depth estimator with a stereo matcher; if the fused mask improves further, it would confirm that the bottleneck is depth quality rather than the fusion model itself.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

5 major / 5 minor

Summary. The paper proposes Dy3DGS-SLAM, a monocular RGB-only 3D Gaussian Splatting SLAM system for dynamic scenes. It fuses an optical-flow-based dynamic mask with a monocular-depth-based mask through a probabilistic fusion step, uses the fused mask in a motion loss for pose estimation, and applies color and depth losses that penalize dynamic Gaussians during mapping. The method is evaluated on TUM RGB-D, BONN RGB-D, and AirDOS-Shibuya, with ATE tracking results reported and rendering shown qualitatively. The central claim is that this is the first RGB-only 3DGS SLAM for dynamic environments and that it outperforms or matches RGB-D-based methods on the tested benchmarks.

Significance. If the claims hold, the paper addresses a genuine gap: extending 3DGS SLAM to dynamic scenes without depth sensors. The experiments use public datasets and compare against several strong baselines, including NICE-SLAM, ESLAM, Co-SLAM, NID-SLAM, and SplaTAM, and the ablation in Table III indicates that fusing optical flow and depth masks improves tracking ATE substantially. There is no apparent circularity in the evaluation. However, the paper's central quantitative claims depend on an underspecified scale factor S_n for monocular depth, and the mapping/rendering claims are not supported by quantitative metrics. The paper also does not release code, and several key components are described only at a high level. These issues make the current version difficult to verify, but the approach is plausible and the weaknesses appear addressable in revision.

major comments (5)
  1. [§III-B, Eqs. (5)-(6)] The scale factor S_n is never defined. The paper states that DepthAnythingV2 provides estimated depth and that S_n provides accurate scale information, but it does not explain how S_n is computed, learned, initialized, or updated. DepthAnythingV2 outputs affine-invariant monocular depth, not metric depth. Without a stated mechanism converting this to metric scale, the scaling of the optical flow in Eq. (5) and the translation normalization in Eq. (6) have no well-defined effect, and the ATE values reported in centimeters cannot be interpreted as metric errors. This is a load-bearing issue because the headline claims of robust tracking and reconstruction depend on recovering metric-scale poses and geometry from monocular RGB. Please specify the computation of S_n, its per-frame dependency, and how scale drift is controlled.
  2. [§III-B, Eq. (6)] The motion loss in Eq. (6) mixes normalized translation terms with a raw rotation difference, and the mask M_ds is applied only to the rotation term. The translation terms are normalized by max(|T·S_n|, ε), while the rotation term (R_hat - R)·M_ds is not normalized or weighted relative to the translation terms, so the loss components have inconsistent scales. Additionally, if dynamic objects should be excluded from tracking, it is unclear why the translation terms are not also masked with M_ds. Please clarify the intended weighting and masking, and specify whether the same mask is applied to both translation and rotation.
  3. [§III-A, Eq. (3)] The Bayesian fusion model is underspecified. Equation (3) states P(D_m, F_m|M(p)) = P(D_m|M(p))·P(F_m|M(p)), but the likelihoods P(D_m|M(p)) and P(F_m|M(p)) and the prior P(M(p)) are never defined. Without these distributions, the posterior in Eq. (4) cannot be computed, and the fusion method is not reproducible. Please provide the actual likelihood models or an explicit algorithmic description of how the depth mask and optical flow mask are combined to produce the thresholded mask in Eq. (4).
  4. [§IV-B and Fig. 2] The paper claims state-of-the-art tracking and rendering in dynamic environments, but only tracking is evaluated quantitatively via ATE. Rendering is assessed only through qualitative mesh images in Fig. 2; no PSNR, SSIM, LPIPS, or geometric reconstruction metrics are reported. This leaves the high-fidelity reconstruction claim unsupported. Please add quantitative rendering and reconstruction comparisons on TUM and BONN against the RGB-D baselines cited in the tables.
  5. [§V-A, Table III] The ablation study in Table III is performed on the AirDOS-Shibuya dataset, while the central tracking and rendering claims are made on TUM and BONN. The table reports an improvement from 7.6 cm to 3.0 cm ATE, but it is not clear whether the same gains hold on the main evaluation sequences. Please include ablations on at least the TUM and BONN sequences used in Tables I and II, or justify why AirDOS-Shibuya is representative.
minor comments (5)
  1. [Table I] The DytanVO row contains the entry '9.05.5', which appears to be a typo for '9.0 5.5'; please reformat all table entries and column headers for readability.
  2. [Eqs. (12)-(13)] The notation in the depth loss is inconsistent: N_d is used as a pixel count in Eq. (12), while D_d is used both as a count and as the depth corresponding to the dynamic mask in Eq. (13). Please define every symbol and keep units consistent.
  3. [§III] Several hyperparameters are not specified, including the K-means cluster count k, the training schedule, the optimizer, and the values of lambda_1, lambda_2, lambda_d, lambda_s, lambda_t, lambda_m, and lambda. This information is needed to reproduce the method.
  4. [§I and §VI] The introduction calls the method 'the first RGB-only 3DGS-SLAM system for dynamic environments,' while the conclusion says 'the first 3DGS-based SLAM method designed for dynamic scenes using monocular RGB input'; please use one consistent statement of novelty.
  5. [§IV-A] The comparison categories in Tables I and II list DytanVO, a learning-based VO method, under 'Traditional SLAM methods'; consider separating learning-based and classical methods for clarity.

Circularity Check

0 steps flagged · score 0.0 of 10

No circularity found: tracking and mapping are evaluated against external benchmarks, and the self-citations are descriptive only.

full rationale

Dy3DGS-SLAM does not derive its headline results from its own outputs. Tracking is assessed with ATE RMSE against ground-truth trajectories from the external TUM RGB-D and BONN RGB-D datasets, and mapping is assessed qualitatively. The fused dynamic mask is computed from optical flow and monocular depth estimates supplied by external networks (a U-Net and DepthAnythingV2), not from the evaluation ground truth or from the method's own rendered geometry. The pose network follows the external TartanVO training design, and the motion and rendering losses use the estimated depth and masks as supervision; none of these quantities is a renamed version of the reported result. The self-citations to DDN-SLAM [9], OVD-SLAM [19], and the survey [7] appear only in the related-work discussion and are not load-bearing for the method's derivation. The per-frame scale factor S_n in Eqs. (5) and (6) is under-specified and could be a correctness risk, but the paper gives no equation or procedure showing that S_n is fitted to the ATE targets or to the ground-truth scale; under-specification is not circularity. No step in the derivation reduces by construction to its own input, so the appropriate circularity score is 0.

Assumptions & free parameters 7 free parameters · 5 assumptions · 0 invented entities

The paper introduces no new physical entities. Its load-bearing assumptions are the conditional independence of optical flow and depth masks, the recoverability of metric scale from monocular depth through an unspecified S_n, and the transferability of pretrained networks. The many unreported hyperparameters are free parameters that directly control the reported performance.

free parameters (7)
  • Mask binarization threshold T = 0.95
    Threshold in Eq. (4) for classifying a pixel as dynamic; set by hand, with no sensitivity analysis.
  • K-means cluster count k = not specified
    Number of moving-object clusters in Eq. (1); the selection method is absent.
  • Tracking loss weights lambda1, lambda2 = not reported
    Weights balancing optical flow, motion segmentation, and pose losses in Eq. (7).
  • Photometric penalty weights lambda_d, lambda_s = not reported
    Weights for dynamic and static pixel regions in Eq. (12).
  • Depth penalty weights lambda_t, lambda_m = not reported
    Weights for dynamic and static depth regions in Eq. (13).
  • Rendering loss weight lambda = 1
    Overall weight in Eq. (14), stated as set to 1.
  • Scale factor S_n = unspecified
    Per-frame scale used to correct monocular depth scale in Eqs. (5) and (6); computation is not described.
assumptions (5)
  • domain assumption Optical flow mask and depth mask are conditionally independent given the motion state M(p).
    Used to factor the posterior in Eq. (3); no justification is given, and both cues depend on image content and motion magnitude.
  • domain assumption Monocular depth estimates can be scaled to metric accuracy by S_n and can constrain pose scale.
    Underpins the scale-corrected motion loss in Eq. (6) and the depth rendering loss in Eq. (13); S_n computation is unspecified.
  • domain assumption K-means clusters in pixel space correspond to distinct moving objects.
    Used to group dynamic pixels in Section III-A before per-object depth search; compactness in image space need not match object identity.
  • domain assumption Pretrained optical flow segmentation, Depth Anything V2, and TartanVO-style networks transfer to the evaluation datasets.
    The paper relies on these pretrained components without reporting fine-tuning details or domain adaptation analysis.
  • domain assumption Dynamic Gaussians can be pruned by setting depth to infinity, and color and depth losses remove the resulting artifacts.
    Central mapping strategy in Section III-C; no analysis is provided for cases where pruning removes static pixels or leaves artifacts.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Dy3DGS-SLAM: Monocular 3D Gaussian Splatting SLAM for Dynamic Environments." pith.science (2026). https://pith.science/paper/XA7XJ75X

@misc{pith2026250605965,
  author       = {Pith},
  title        = {Pith review of: Dy3DGS-SLAM: Monocular 3D Gaussian Splatting SLAM for Dynamic Environments},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/XA7XJ75X}},
  note         = {Machine review of arXiv:2506.05965}
}
read the original abstract

Current Simultaneous Localization and Mapping (SLAM) methods based on Neural Radiance Fields (NeRF) or 3D Gaussian Splatting excel in reconstructing static 3D scenes but struggle with tracking and reconstruction in dynamic environments, such as real-world scenes with moving elements. Existing NeRF-based SLAM approaches addressing dynamic challenges typically rely on RGB-D inputs, with few methods accommodating pure RGB input. To overcome these limitations, we propose Dy3DGS-SLAM, the first 3D Gaussian Splatting (3DGS) SLAM method for dynamic scenes using monocular RGB input. To address dynamic interference, we fuse optical flow masks and depth masks through a probabilistic model to obtain a fused dynamic mask. With only a single network iteration, this can constrain tracking scales and refine rendered geometry. Based on the fused dynamic mask, we designed a novel motion loss to constrain the pose estimation network for tracking. In mapping, we use the rendering loss of dynamic pixels, color, and depth to eliminate transient interference and occlusion caused by dynamic objects. Experimental results demonstrate that Dy3DGS-SLAM achieves state-of-the-art tracking and rendering in dynamic environments, outperforming or matching existing RGB-D methods.

Figures

Figures reproduced from arXiv: 2506.05965 by the authors.

Figure 1
Figure 1. Pipeline of Our Network: Our system workflow consists of two main threads: tracking and mapping. In the tracking thread, we use a segmentation optical flow network and a depth estimation network to generate the estimated motion optical flow mask and depth map mask. By applying a conditional probability approach, we create a fused mask Mˆ . This fused mask is subsequently input into the pose estimation network to det… view at source ↗
Figure 2
Figure 2. Visual comparison of the reconstructed meshes on the BONN and TUM RGB-D datasets. Our results are more [PITH_FULL_IMAGE:figures/full_fig_p005_2.png] view at source ↗
Figure 3
Figure 3. Evaluation of Tracking Network Loss Methods [PITH_FULL_IMAGE:figures/full_fig_p006_3.png] view at source ↗

Discussion (0). Sign in to comment.

Forward citations

Cited by 2 Pith papers

Reviewed papers in the Pith corpus that reference this work. Sorted by Pith novelty score. Full citation record

  1. CubifyGS: Object-Centric 3D Gaussian Splatting for Lifelong Dynamic Scene Maintenance

    cs.RO 2026-06 unverdicted novelty 6.0 of 10

    CubifyGS maintains 3DGS maps under rigid object rearrangements via reusable Gaussian assets, explicit prune/retrieve/align operations, and event-triggered local optimization, outperforming baselines on a new dynamic b...

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

    cs.CV 2026-02 conditional novelty 6.0 of 10

    RU4D-SLAM reweights per-pixel uncertainty to map static and dynamic Gaussians, handling motion blur and exposure shifts, beating previous Gaussian-SLAM on TUM, Bonn, and Wild-SLAM.

Reference graph

Works this paper leans on

39 extracted references · 25 canonical work pages · cited by 2 Pith papers

  1. [1]

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

    B. Mildenhall, P. P. Srinivasan, M. Tancik, J. T. Barron, R. Ramamoor- thi, and R. Ng, “Nerf: Representing scenes as neural radiance fields for view synthesis,”Communications of the ACM, vol. 65, no. 1, pp. 99–106, 2021

    2021

  2. [2]

    3d gaussian splatting for real-time radiance field rendering

    B. Kerbl, G. Kopanas, T. Leimk ¨uhler, and G. Drettakis, “3d gaussian splatting for real-time radiance field rendering.”ACM Transactions on Graphics (TOG), vol. 42, no. 4, pp. 139–1, 2023

    work page 2023

  3. [3]

    imap: Implicit map- ping and positioning in real-time,

    E. Sucar, S. Liu, J. Ortiz, and A. J. Davison, “imap: Implicit map- ping and positioning in real-time,” inProceedings of the IEEE/CVF International Conference on Computer Vision, 2021, pp. 6229–6238

    work page 2021

  4. [4]

    Nice-slam: Neural implicit scalable encoding for slam,

    Z. Zhu, S. Peng, V . Larsson, W. Xu, H. Bao, Z. Cui, M. R. Oswald, and M. Pollefeys, “Nice-slam: Neural implicit scalable encoding for slam,” inProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2022, pp. 12 786–12 796

    work page 2022

  5. [5]

    Point- slam: Dense neural point cloud-based slam,

    E. Sandstr ¨om, Y . Li, L. Van Gool, and M. R. Oswald, “Point- slam: Dense neural point cloud-based slam,” inProceedings of the IEEE/CVF International Conference on Computer Vision, 2023, pp. 18 433–18 444

    2023

  6. [6]

    How nerfs and 3d gaussian splatting are reshaping slam: a survey,

    F. Tosi, Y . Zhang, Z. Gong, E. Sandstr¨om, S. Mattoccia, M. R. Oswald, and M. Poggi, “How nerfs and 3d gaussian splatting are reshaping slam: a survey,”arXiv preprint arXiv:2402.13255, vol. 4, 2024

    arXiv 2024

  7. [7]

    Evaluating Modern Approaches in 3D Scene Reconstruction: NeRF vs Gaussian-Based Methods

    Y . Zhou, Z. Zeng, A. Chen, X. Zhou, H. Ni, S. Zhang, P. Li, L. Liu, M. Zheng, and X. Chen, “Evaluating modern approaches in 3d scene reconstruction: Nerf vs gaussian-based methods,”arXiv preprint arXiv:2408.04268, 2024

    work page Pith review arXiv 2024

  8. [8]

    Dn-slam: A visual slam with orb features and nerf mapping in dynamic environments,

    C. Ruan, Q. Zang, K. Zhang, and K. Huang, “Dn-slam: A visual slam with orb features and nerf mapping in dynamic environments,”IEEE Sensors Journal, vol. 24, no. 4, pp. 5279–5287, 2024

    work page 2024

Show all 39 references
  1. [9]

    Ddn-slam: Real-time dense dynamic neural implicit slam with joint semantic encoding,

    M. Li, Y . Zhou, G. Jiang, T. Deng, Y . Wang, and H. Wang, “Ddn-slam: Real-time dense dynamic neural implicit slam with joint semantic encoding,”arXiv preprint arXiv:2401.01545, 2024

    2024 arXiv

  2. [10]

    Nid-slam: Neural implicit representation-based rgb-d slam in dynamic environments,

    Z. Xu, J. Niu, Q. Li, T. Ren, and C. Chen, “Nid-slam: Neural implicit representation-based rgb-d slam in dynamic environments,” arXiv preprint arXiv:2401.01189, 2024

    2024 arXiv

  3. [11]

    Rodyn-slam: Robust dynamic dense rgb-d slam with neural radiance fields,

    H. Jiang, Y . Xu, K. Li, J. Feng, and L. Zhang, “Rodyn-slam: Robust dynamic dense rgb-d slam with neural radiance fields,”IEEE Robotics and Automation Letters, vol. 9, no. 9, pp. 7509–7516, 2024

    2024

  4. [12]

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

    N. Keetha, J. Karhade, K. M. Jatavallabhula, G. Yang, S. Scherer, D. Ramanan, and J. Luiten, “Splatam: Splat track & map 3d gaussians for dense rgb-d slam,” inProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2024, pp. 21 357–21 366

    2024

  5. [13]

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

    H. Huang, L. Li, H. Cheng, and S.-K. Yeung, “Photo-slam: Real-time simultaneous localization and photorealistic mapping for monocular stereo and rgb-d cameras,” inProceedings of the IEEE/CVF Confer- ence on Computer Vision and Pattern Recognition, 2024, pp. 21 584– 21 593

    2024

  6. [14]

    Gaussian splatting slam,

    H. Matsuki, R. Murai, P. H. Kelly, and A. J. Davison, “Gaussian splatting slam,” inProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2024, pp. 18 039–18 048

    2024

  7. [15]

    Demonstration of chromocloth: Re-programmable multi-color textures through flexible and portable light source,

    Y . Zhu, C. Honnet, Y . Kang, J. Zhu, A. J. Zheng, K. Heinz, G. Tang, L. Musk, M. Wessely, and S. Mueller, “Demonstration of chromocloth: Re-programmable multi-color textures through flexible and portable light source,” inAdjunct Proceedings of the 36th Annual ACM Symposium on...

    2023

  8. [16]

    Going blank comfortably: Positioning monocular head- worn displays when they are inactive,

    Y . Song, P. Arora, R. Singh, S. T. Varadharajan, M. Haynes, and T. Starner, “Going blank comfortably: Positioning monocular head- worn displays when they are inactive,” inProceedings of the ACM International Symposium on Wearable Computers, 2023, pp. 114–118

    2023

  9. [17]

    Looking from a different angle: Placing head-worn displays near the nose,

    Y . Song, P. Arora, S. T. Varadharajan, R. Singh, M. Haynes, and T. Starner, “Looking from a different angle: Placing head-worn displays near the nose,” inProceedings of the Augmented Humans International Conference, 2024, pp. 28–45

    2024

  10. [18]

    Ds-slam: A semantic visual slam towards dynamic environments,

    C. Yu, Z. Liu, X.-J. Liu, F. Xie, Y . Yang, Q. Wei, and Q. Fei, “Ds-slam: A semantic visual slam towards dynamic environments,” inIEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), 2018, pp. 1168–1174

    2018

  11. [19]

    Ovd-slam: An online visual slam for dynamic environments,

    J. He, M. Li, Y . Wang, and H. Wang, “Ovd-slam: An online visual slam for dynamic environments,”IEEE Sensors Journal, vol. 23, no. 12, pp. 13 210–13 219, 2023

    2023

  12. [20]

    Sg-slam: A real- time rgb-d visual slam toward dynamic scenes with semantic and geometric information,

    S. Cheng, C. Sun, S. Zhang, and D. Zhang, “Sg-slam: A real- time rgb-d visual slam toward dynamic scenes with semantic and geometric information,”IEEE Transactions on Instrumentation and Measurement, vol. 72, pp. 1–12, 2022

    2022

  13. [21]

    Flowfusion: Dynamic dense rgb-d slam based on optical flow,

    T. Zhang, H. Zhang, Y . Li, Y . Nakamura, and L. Zhang, “Flowfusion: Dynamic dense rgb-d slam based on optical flow,” inIEEE Inter- national Conference on Robotics and Automation (ICRA), 2020, pp. 7322–7328

    2020

  14. [22]

    Deflowslam: Self-supervised scene motion decomposition for dynamic dense slam,

    W. Ye, X. Yu, X. Lan, Y . Ming, J. Li, H. Bao, Z. Cui, and G. Zhang, “Deflowslam: Self-supervised scene motion decomposition for dynamic dense slam,”arXiv preprint arXiv:2207.08794, 2022

    2022 arXiv

  15. [23]

    Dytanvo: Joint refinement of visual odometry and motion segmentation in dynamic environ- ments,

    S. Shen, Y . Cai, W. Wang, and S. Scherer, “Dytanvo: Joint refinement of visual odometry and motion segmentation in dynamic environ- ments,” inIEEE International Conference on Robotics and Automation (ICRA), 2023, pp. 4048–4055

    2023

  16. [24]

    Instant neural graphics primitives with a multiresolution hash encoding,

    T. M ¨uller, A. Evans, C. Schied, and A. Keller, “Instant neural graphics primitives with a multiresolution hash encoding,”ACM Transactions on Graphics (TOG), vol. 41, no. 4, pp. 1–15, 2022

    2022

  17. [25]

    U-net: Convolutional networks for biomedical image segmentation,

    O. Ronneberger, P. Fischer, and T. Brox, “U-net: Convolutional networks for biomedical image segmentation,” inMedical Image Computing and Computer-Assisted Intervention (MICCAI), 2015

    2015

  18. [26]

    Depth anything v2,

    L. Yang, B. Kang, Z. Huang, Z. Zhao, X. Xu, J. Feng, and H. Zhao, “Depth anything v2,”arXiv preprint arXiv:2406.09414, 2024

    2024 arXiv

  19. [27]

    Tartanvo: A generalizable learning- based vo,

    W. Wang, Y . Hu, and S. Scherer, “Tartanvo: A generalizable learning- based vo,” inConference on Robot Learning. PMLR, 2021, pp. 1761–1772

    2021

  20. [28]

    Deep residual learning for image recognition,

    K. He, X. Zhang, S. Ren, and J. Sun, “Deep residual learning for image recognition,” inProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2016, pp. 770–778

    2016

  21. [29]

    Point-based neural rendering with per-view optimization,

    G. Kopanas, J. Philip, T. Leimk ¨uhler, and G. Drettakis, “Point-based neural rendering with per-view optimization,” inComputer Graphics F orum, vol. 40, no. 4. Wiley Online Library, 2021, pp. 29–43

    2021

  22. [30]

    Differentiable surface splatting for point-based geometry processing,

    W. Yifan, F. Serena, S. Wu, C. ¨Oztireli, and O. Sorkine-Hornung, “Differentiable surface splatting for point-based geometry processing,” ACM Transactions on Graphics (TOG), vol. 38, no. 6, pp. 1–14, 2019

    2019

  23. [31]

    A benchmark for the evaluation of rgb-d slam systems,

    J. Sturm, N. Engelhard, F. Endres, W. Burgard, and D. Cremers, “A benchmark for the evaluation of rgb-d slam systems,” inIEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), 2012, pp. 573–580

    2012

  24. [32]

    Airdos: Dynamic slam benefits from articulated objects,

    Y . Qiu, C. Wang, W. Wang, M. Henein, and S. Scherer, “Airdos: Dynamic slam benefits from articulated objects,” inIEEE International Conference on Robotics and Automation (ICRA), 2022, pp. 8047– 8053

    2022

  25. [33]

    Refusion: 3d reconstruction in dynamic environments for rgb-d cameras exploiting residuals,

    E. Palazzolo, J. Behley, P. Lottes, P. Gigu `ere, and C. Stachniss, “Refusion: 3d reconstruction in dynamic environments for rgb-d cameras exploiting residuals,” inIEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), 2019, pp. 7855–7862

    2019

  26. [34]

    Orb-slam3: An accurate open-source library for visual, visual–inertial, and multimap slam,

    C. Campos, R. Elvira, J. J. G. Rodr ´ıguez, J. M. Montiel, and J. D. Tard ´os, “Orb-slam3: An accurate open-source library for visual, visual–inertial, and multimap slam,”IEEE Transactions on Robotics, vol. 37, no. 6, pp. 1874–1890, 2021

    2021

  27. [35]

    Droid-slam: Deep visual slam for monocular, stereo, and rgb-d cameras,

    Z. Teed and J. Deng, “Droid-slam: Deep visual slam for monocular, stereo, and rgb-d cameras,”Advances in Neural Information Process- ing Systems, vol. 34, pp. 16 558–16 569, 2021

    2021

  28. [36]

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

    M. M. Johari, C. Carta, and F. Fleuret, “Eslam: Efficient dense slam system based on hybrid representation of signed distance fields,” in Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2023, pp. 17 408–17 419

    2023

  29. [37]

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

    H. Wang, J. Wang, and L. Agapito, “Co-slam: Joint coordinate and sparse parametric encodings for neural real-time slam,” inProceed- ings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2023, pp. 13 293–13 302

    2023

  30. [38]

    Dynaslam: Tracking, mapping, and inpainting in dynamic scenes,

    B. Bescos, J. M. F ´acil, J. Civera, and J. Neira, “Dynaslam: Tracking, mapping, and inpainting in dynamic scenes,”IEEE Robotics and Automation Letters, vol. 3, no. 4, pp. 4076–4083, 2018

    2018

  31. [39]

    Dense visual slam for rgb-d cameras,

    C. Kerl, J. Sturm, and D. Cremers, “Dense visual slam for rgb-d cameras,” inIEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), 2013, pp. 2100–2106

    2013

Pith tools

Reviewed August 7, 2026 · model on record in the stance chip above.