Pith. sign in

REVIEW 4 major objections 5 minor 33 references

GARAD-SLAM: 3D GAussian splatting for Real-time Anti Dynamic SLAM

T0 review · 4 major / 5 minor · reviewed 2026-08-09 · deepseek-v4-flash

Pith's one-line read GARAD-SLAM claims that labeling 3D Gaussians as static or dynamic, then mapping those labels back to front-end features, lets a 3DGS SLAM system track and reconstruct dynamic scenes in real time without rendering artifacts.

desk verdict Plausible first 3DGS dynamic-SLAM system with a real bottleneck addressed, but the core equations are under-specified to the point of non-reproducibility; worth a referee only with code and full details. read the letter →

arxiv 2502.03228 v2 pith:5FWYFG2C submitted 2025-02-05 cs.RO cs.CV

classification cs.ROcs.CV
keywords 3DGaussianSplattingSLAMdynamicscenesconditionalrandomfieldpyramidnetworkopticalflowverificationRGB-Dreal-timerendering
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

GARAD-SLAM sets out to show that a 3D Gaussian Splatting SLAM system can handle rooms with moving people by classifying the Gaussians themselves as static or dynamic, instead of running a separate pixel-level semantic segmentation. The labels are computed on the map side with a conditional random field, then mapped back to the front-end feature points, so tracking and mapping can correct each other rather than being loosely coupled. The paper argues that this two-way coupling removes the ghost artifacts that dynamic objects leave in 3DGS reconstructions while also preventing the irreversible deletion of static structure. On the TUM and BONN dynamic RGB-D benchmarks, the system reports absolute trajectory error competitive with or better than existing dynamic SLAM methods, with rendering that shows fewer artifacts.

What carries the argument

The load-bearing object is the Tagged Gaussian: a 3D Gaussian primitive augmented with an ORB feature point, a rotation, a density, and spherical-harmonic color descriptors, so the same primitive anchors both the map and the front-end feature tracker. The decision machinery is a fully connected CRF with a Gaussian-mixture unary potential (Eq. 2) and two pairwise kernels (appearance/observation and position), plus a chi-square optical-flow test that reclassifies suspicious labels. The dynamic-rendering penalty $L_{\mathrm{dyn}} = \sum_{i \in G_D} \alpha_i^2$ is what lets the map degrade moving Gaussians gradually instead of pruning them instantly.

What would settle it

Run GARAD-SLAM on a dynamic RGB-D sequence in which a person walks slowly along the camera's epipolar line for many frames, then compare ATE and render quality against the same sequence with fast lateral motion. If the four CRF cues cannot separate the moving body from static structure, the dynamic Gaussians stay in the map and the reported anti-dynamic advantage should shrink or vanish; the paper's own ablation (w/o Flow) indicates that without the optical-flow recovery the TUM ATE degrades from 0.0194 m to 0.0232 m, so the recovery stage is the first place a failure would appear.

Watch

Extended reading notes

Core claim

The central claim is that dynamic/static segmentation belongs on the Gaussians, not on the pixels. GARAD-SLAM constructs a fully connected CRF over tagged Gaussians, whose unary potentials come from a Gaussian mixture model over four cues—reprojection error, depth variation, observation count, and epipolar distance—and whose pairwise kernels enforce spatial smoothness. Once a Gaussian is labeled dynamic, the label is propagated through a Gaussian pyramid network to the ORB feature points that the Gaussian carries, so the front-end pose optimizer simply discards dynamic matches. A sparse optical-flow verification step then recovers static Gaussians that the CRF mislabeled, and a dynamic-label rendering penalty suppresses the contribution of dynamic Gaussians during optimization while a time-window retention rule delays their deletion. The claimed result is a tightly coupled tracking-and-mapping loop in which pose accuracy and map cleanliness improve together.

Load-bearing premise

The system assumes a moving object can always be told apart from static structure by four measurable cues—reprojection error, depth change, observation count, and epipolar distance—so that a Gaussian mixture cleanly separates the two populations; when an object moves slowly, is observed for a long time, or travels along the epipolar line, that separation fails and the whole dynamic-label pipeline loses its foundation.

Editorial extensions

If this is right

  • Dynamic objects can be handled without pixel-level semantic segmentation, which removes a costly front-end stage and lets 3DGS SLAM stay real-time.
  • Tracking benefits directly from map-side labels, so pose drift caused by moving features is reduced on dynamic benchmarks.
  • The time-window retention and quadratic opacity penalty mean misclassified static Gaussians can be recovered instead of being permanently pruned.
  • The same tagged-Gaussian representation allows the map to be built as a static scene even while people move through it, which is the requirement for AR/VR and robot navigation in human environments.

Reading between the lines

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

  • Because the CRF unary model is cue-based rather than semantic, the system should generalize to any moving object regardless of class, but it should also fail in the same way for slow, epipolar-aligned motion; that is a testable prediction the paper does not run.
  • The Gaussian pyramid supervision could be reused to propagate finer-grained labels, such as per-object instance ids, turning the dynamic mask into a moving-object tracker with little extra machinery.
  • The static-flow Gaussian assumption in Eq. 8 is likely the first component to break under rapid rotation, so a rotation-compensated flow model is a natural extension for handheld cameras.
  • If the rendering penalties are strong enough, the method may double as a video inpainting tool: the retained dynamic Gaussians can be re-rendered later to reconstruct the moving person, not just remove them.
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

4 major / 5 minor

Summary. The paper proposes GARAD-SLAM, a 3D Gaussian Splatting (3DGS)-based RGB-D SLAM system for dynamic scenes. In tracking, ORB feature points are associated with 'Tagged Gaussians'; a fully connected CRF on Gaussians produces dynamic/static labels, which are mapped back to front-end feature points via a 'Gaussian pyramid network.' A sparse optical-flow chi-square test is used to recover incorrectly labeled static points. The map is optimized with a photometric-SSIM loss plus a dynamic-label penalty that suppresses dynamic Gaussians instead of immediately deleting them. The system is evaluated on TUM and BONN dynamic sequences, reporting ATE/STD comparable to DDN-SLAM and better than ORB-SLAM3, Photo-SLAM, SplaTAM, and NID-SLAM, together with an ablation study and runtime comparison.

Significance. The core idea of deriving dynamic labels directly in Gaussian space and feeding them back into feature-based tracking is a useful extension of 3DGS-based SLAM, and the reported tracking numbers are competitive. The paper evaluates on external benchmarks, provides per-sequence ATE/STD, and includes ablations and runtime, which are strengths. The formulation is not circular by construction: labels come from measured features rather than from the pose output being evaluated. However, the method is currently under-specified in ways that prevent verification: the CRF/GMM unary model is not a well-defined likelihood, the 'Gaussian pyramid network' is described only narratively, and many free parameters are not reported. These gaps block independent reproduction and therefore also the assessment of whether the reported gains come from the proposed mechanism or from hidden implementation choices.

major comments (4)
  1. [III-A, Eq. (2)] Equation (2) is not a well-defined likelihood. The text defines four Gaussian models N^α_i, N^β_i, N^γ_i, N^δ_i over reprojection error, depth variation, observation count, and epipolar distance, but Eq. (2) writes P_static_i = Σ_{k=1}^4 π_k N(x|μ_k, σ_k^2), which is a mixture over one scalar x rather than a combination of the four per-feature Gaussian models. No feature vector, normalization, covariance structure, mixing weights, or fitting procedure is provided, so the unary potential in Eq. (3) cannot be computed from the stated inputs. This is load-bearing because Eq. (7), the label back-mapping, Eq. (10), and Eq. (11) all depend on these labels. Please specify the four-dimensional feature vector, the per-feature Gaussian parameters, and the estimation or calibration procedure.
  2. [III-A, Eq. (14)] The 'Gaussian pyramid network' that is central to the contributions (Sec. I, contribution 3) is never defined as a network. Eq. (14) is only a multiresolution loss between rendered images and pyramid-smoothed ground truth images; it does not specify an architecture, the layer at which labels are assigned, how Gaussian labels are mapped back to ORB feature points, or how the network is trained. Without this, the claimed back-propagation of dynamic labels to the front end and the label updates used in Sec. III-C are not reproducible. Please provide the network definition and the exact back-projection rule.
  3. [III-B, Eqs. (8)-(9)] The optical-flow verification assumes that optical-flow vectors of 'static points' follow a single Gaussian distribution. This is not justified for general camera motion: under rotation, static flow is approximately determined by angular velocity and is not an isotropic Gaussian, and with parallax it varies with depth. The selection rule for static points is also unspecified ('we select a certain number of static points' with no criterion), and if the selection already uses the CRF labels, the verification is partially circular. Please state how the static point set is chosen and provide a calibration or validation of the Gaussian/chi-square model.
  4. [IV-A, Eqs. (5)-(7)] The free parameters of the method are under-reported. The text fixes only λ_p−ssim=0.8, λ_dyn=0.2, λ=0.2, and the pyramid level n=2; it does not state the GMM parameters in Eq. (2), the kernel bandwidths σ_α, σ_γ, σ_P, σ_p in Eqs. (5)-(6), the retention window n in Eq. (7), the chi-square threshold in Eq. (9), or any pruning thresholds. Since the ATE results in Tables I-III depend on these parameters, please report all values and, ideally, a sensitivity analysis.
minor comments (5)
  1. [III-C, Eq. (12)] The notation uses Tgt in Eq. (12) while the text and surrounding equations use Igt; also SSIM(I_r - I_gt) is unusual, and the intended term is likely SSIM(I_r, I_gt).
  2. [Fig. 2] Figure 2 is difficult to read: it contains placeholder '??' symbols and the text is very small; please replace it with a readable version.
  3. [Table III] The column label 'w/o Full' is confusing because the text says it refers to the full method (CRF segmentation plus sparse optical flow verification); rename it to 'Full' or 'Ours'.
  4. [IV-A, Implementation details] The hardware is listed as an 'RTX 4080 Ti', which is not a standard product name; please correct to the actual GPU model used.
  5. [I, Introduction] The introduction claims 'state-of-the-art results in rendering accuracy', but the experiments only provide qualitative rendering comparisons because no ground-truth reconstructions are available; please either add a quantitative rendering metric with appropriate masks or soften the claim.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the tracking and dynamic-segmentation claims are evaluated against external benchmarks, and the self-citations (OVD-SLAM, DDN-SLAM) are background/baseline rather than load-bearing derivation steps.

full rationale

GARAD-SLAM's central claims are evaluated on external TUM and BONN benchmarks with ground-truth ATE, so the tracking result is not constructed from the method's own outputs. The dynamic/static labels are produced by CRF inference over measured per-Gaussian features (reprojection error, depth variation, observation count, epipolar distance) and then refined by sparse optical-flow chi-square testing; the pose is optimized over the resulting static set. This is an iterative estimation loop, not a circular derivation: the label is not defined as the pose, and the pose is not defined as the label. The self-citations (OVD-SLAM [9], DDN-SLAM [17]) appear only as related work and as comparison baselines; no load-bearing theorem or design constraint is imported from them, and the paper does not claim a uniqueness result based on its own prior work. The main formal weakness is that Eq. (2) is under-specified: the four Gaussian models N^alpha..N^delta are described but the displayed mixture uses a single scalar x with no feature vector, normalization, or fitting procedure, and CRF/GMM parameters and thresholds are unreported. That is a reproducibility/correctness gap, not a circular reduction, because no fitted parameter is renamed as a prediction and no equation reduces to its own input by construction.

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

The method is built on several unstated parameter values and domain assumptions: CRF/GMM parameters, kernel bandwidths, retention window, pruning thresholds, and static-point selection for optical flow. The initial ten-frame static-scene assumption and the four statistical assumptions about dynamic Gaussians are load-bearing but not validated. No new physical entities are introduced; 'Tagged Gaussians' is an augmented data structure, not a new physical postulate.

free parameters (6)
  • CRF unary GMM parameters (πk, μk, σk) = not reported
    These parameters in Eq. (2) define the static likelihood for reprojection error, depth variation, observation count, and epipolar distance; no values or fitting procedure are given.
  • Pairwise kernel bandwidths (σ_α, σ_γ, σ_P, σ_p) = not reported
    Appear in Eqs. (5) and (6) and control CRF label smoothness. The values are not stated, so the segmentation result depends on unstated choices.
  • Dynamic label retention window n = not reported
    Eq. (7) delays deletion until a Gaussian is labeled dynamic for n consecutive frames; n is never specified.
  • Loss weights in Eqs. (11) and (12) = λ_p-ssim=0.8, λ_dyn=0.2, λ=0.2
    Set in Sec IV-A by hand; they affect the balance between photometric loss, SSIM, and the dynamic penalty.
  • Gaussian pruning thresholds = not reported
    Sec III-C prunes Gaussians with low opacity and large size, but the thresholds are unspecified.
  • Static-point set for optical flow model = not reported
    Sec III-B requires selecting static points to fit a Gaussian distribution of flow vectors; the selection rule and sample size are unspecified.
assumptions (6)
  • domain assumption The first ten frames are assumed to be a static scene for initialization.
    Sec III-A states the system assumes a static scene during the first ten frames. If dynamic objects appear during initialization, the Tagged Gaussian map and CRF baseline could be corrupted.
  • domain assumption Dynamic Gaussians have larger reprojection error, larger depth variation, fewer observations, and larger epipolar distance than static Gaussians.
    These four assumptions justify the unary potential in Eq. (2). The paper provides no quantitative support or validation for them.
  • domain assumption Optical flow vectors of static feature points follow a multivariate Gaussian distribution.
    The chi-square test in Eqs. (8) and (9) relies on this assumption. Under camera rotation or parallax, static-point flow is not necessarily Gaussian.
  • standard math Fully connected CRF inference with Gaussian edge potentials is a valid tool for labeling Gaussians.
    Sec III-A invokes the CRF framework of Krähenbühl and Koltun (2011). This is a standard background method.
  • domain assumption The rendering loss with the dynamic alpha penalty drives the optimization to a clean static reconstruction.
    Eq. (11) assumes that minimizing photometric loss plus the alpha penalty over dynamic Gaussians yields correct geometry and appearance. No convergence or identifiability analysis is given.
  • domain assumption Lucas-Kanade optical flow is accurate for matched keypoints.
    The flow vectors Vk feed the Gaussian model in Eq. (8). Large motion, occlusion, or lighting changes can violate this assumption.

how reviews work

0 comments
Cite this review

Pith. "Pith review of GARAD-SLAM: 3D GAussian splatting for Real-time Anti Dynamic SLAM." pith.science (2026). https://pith.science/paper/5FWYFG2C

@misc{pith2026250203228,
  author       = {Pith},
  title        = {Pith review of: GARAD-SLAM: 3D GAussian splatting for Real-time Anti Dynamic SLAM},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/5FWYFG2C}},
  note         = {Machine review of arXiv:2502.03228}
}
read the original abstract

The 3D Gaussian Splatting (3DGS)-based SLAM system has garnered widespread attention due to its excellent performance in real-time high-fidelity rendering. However, in real-world environments with dynamic objects, existing 3DGS-based SLAM systems often face mapping errors and tracking drift issues. To address these problems, we propose GARAD-SLAM, a real-time 3DGS-based SLAM system tailored for dynamic scenes. In terms of tracking, unlike traditional methods, we directly perform dynamic segmentation on Gaussians and map them back to the front-end to obtain dynamic point labels through a Gaussian pyramid network, achieving precise dynamic removal and robust tracking. For mapping, we impose rendering penalties on dynamically labeled Gaussians, which are updated through the network, to avoid irreversible erroneous removal caused by simple pruning. Our results on real-world datasets demonstrate that our method is competitive in tracking compared to baseline methods, generating fewer artifacts and higher-quality reconstructions in rendering.

Figures

Figures reproduced from arXiv: 2502.03228 by the authors.

Figure 1
Figure 1. Rendering result in BONN datasets. The reconstruc [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. Frame of GARAD-SLAM. Given a series of RGB-D frames, we simultaneously construct the gaussian map and [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗
Figure 3
Figure 3. ATE for ORB-SLAM3 and GARAD-SLAM on some sequences of TUM and BONN datasets. [PITH_FULL_IMAGE:figures/full_fig_p005_3.png] view at source ↗
Figures from the paper (2 more)
Figure 4
Figure 4. Figure 4: Gaussian map of TUM RGB-D and BONN datasets. [PITH_FULL_IMAGE:figures/full_fig_p006_4.png]
Figure 5
Figure 5. Figure 5: Comparison of rendering quality. D. Ablation Study In this section, we conduct various experiments to verify the effectiveness of our method. w/o CRF refers to the TABLE III: ABLATION STUDY RESULTS ON TUM AND BONN DATASETS. THE METRIC UNIT IS [m] Dataset Metric w/o CRF…

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

33 extracted references · 21 canonical work pages

  1. [1]

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

    C. Yan, D. Qu, D. Xu, B. Zhao, Z. Wang, D. Wang, and X. Li, “Gs- slam: Dense visual slam with 3d gaussian splatting,” in Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recog- nition, 2024, pp. 19 595–19 604

  2. [2]

    Gaussian splatting slam,

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

  3. [3]

    Gaussian-slam: Photo-realistic dense slam with gaussian splatting,

    V . Yugay, Y . Li, T. Gevers, and M. R. Oswald, “Gaussian-slam: Photo-realistic dense slam with gaussian splatting,” arXiv preprint arXiv:2312.10070, 2023

  4. [4]

    Compact 3d gaussian splatting for dense visual slam,

    T. Deng, Y . Chen, L. Zhang, J. Yang, S. Yuan, D. Wang, and W. Chen, “Compact 3d gaussian splatting for dense visual slam,” arXiv preprint arXiv:2403.11247, 2024

  5. [5]

    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,” in Proceedings of the IEEE/CVF Confer- ence on Computer Vision and Pattern Recognition , 2024, pp. 21 584– 21 593

  6. [6]

    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,” in Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , 2024, pp. 21 357–21 366

  7. [7]

    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

  8. [8]

    Dynamic-slam: Semantic monocular visual localization and mapping based on deep learning in dynamic environment,

    L. Xiao, J. Wang, X. Qiu, Z. Rong, and X. Zou, “Dynamic-slam: Semantic monocular visual localization and mapping based on deep learning in dynamic environment,” Robotics and Autonomous Systems, vol. 117, pp. 1–16, 2019

Show all 33 references
  1. [9]

    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

  2. [10]

    Rds-slam: Real-time dynamic slam using semantic segmentation methods,

    Y . Liu and J. Miura, “Rds-slam: Real-time dynamic slam using semantic segmentation methods,” IEEE Access , vol. 9, pp. 23 772– 23 785, 2021

  3. [11]

    Cfp-slam: A real-time visual slam based on coarse-to-fine probability in dynamic environments,

    X. Hu, Y . Zhang, Z. Cao, R. Ma, Y . Wu, Z. Deng, and W. Sun, “Cfp-slam: A real-time visual slam based on coarse-to-fine probability in dynamic environments,” in IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS) , 2022, pp. 4399–4406

  4. [12]

    Accurate dynamic slam using crf-based long-term consistency,

    Z.-J. Du, S.-S. Huang, T.-J. Mu, Q. Zhao, R. R. Martin, and K. Xu, “Accurate dynamic slam using crf-based long-term consistency,” IEEE Transactions on Visualization and Computer Graphics , vol. 28, no. 4, pp. 1745–1757, 2020

  5. [13]

    Clustervo: Clustering moving instances and estimating visual odometry for self and sur- roundings,

    J. Huang, S. Yang, T.-J. Mu, and S.-M. Hu, “Clustervo: Clustering moving instances and estimating visual odometry for self and sur- roundings,” in Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , 2020, pp. 2168–2177

  6. [14]

    Rgb-d slam in dynamic environments using point correlations,

    W. Dai, Y . Zhang, P. Li, Z. Fang, and S. Scherer, “Rgb-d slam in dynamic environments using point correlations,” IEEE Transactions on Pattern Analysis and Machine Intelligence, vol. 44, no. 1, pp. 373– 389, 2020

  7. [15]

    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

  8. [16]

    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

  9. [17]

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

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

  10. [18]

    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

  11. [19]

    Efficient inference in fully connected crfs with gaussian edge potentials,

    P. Kr ¨ahenb¨uhl and V . Koltun, “Efficient inference in fully connected crfs with gaussian edge potentials,” Advances in Neural Information Processing Systems, vol. 24, 2011

  12. [20]

    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

  13. [21]

    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, vol. 42, no. 4, pp. 1–14, 2023

  14. [22]

    Plgslam: Progressive neural scene represenation with local to global bundle adjustment,

    T. Deng, G. Shen, T. Qin, J. Wang, W. Zhao, J. Wang, D. Wang, and W. Chen, “Plgslam: Progressive neural scene represenation with local to global bundle adjustment,” in Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , 2024, pp. 19 657–19 666

  15. [23]

    Neslam: Neural implicit mapping and self-supervised feature tracking with depth completion and denoising,

    T. Deng, Y . Wang, H. Xie, H. Wang, J. Wang, D. Wang, and W. Chen, “Neslam: Neural implicit mapping and self-supervised feature tracking with depth completion and denoising,” arXiv preprint arXiv:2403.20034, 2024

  16. [24]

    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,” in Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2022, pp. 12 786–12 796

  17. [25]

    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

  18. [26]

    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,” in Proceed- ings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2023, pp. 13 293–13 302

  19. [27]

    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,” in Proceedings of the IEEE/CVF International Conference on Computer Vision , 2021, pp. 6229–6238

  20. [28]

    Dynam-slam: An accurate, robust stereo visual-inertial slam method in dynamic environ- ments,

    H. Yin, S. Li, Y . Tao, J. Guo, and B. Huang, “Dynam-slam: An accurate, robust stereo visual-inertial slam method in dynamic environ- ments,” IEEE Transactions on Robotics , vol. 39, no. 1, pp. 289–308, 2022

  21. [29]

    Blitz-slam: A semantic slam in dynamic environments,

    Y . Fan, Q. Zhang, Y . Tang, S. Liu, and H. Han, “Blitz-slam: A semantic slam in dynamic environments,” Pattern Recognition , vol. 121, p. 108225, 2022

  22. [30]

    Dp-slam: A visual slam with moving probability towards dynamic environments,

    A. Li, J. Wang, M. Xu, and Z. Chen, “Dp-slam: A visual slam with moving probability towards dynamic environments,” Information Sciences, vol. 556, pp. 128–142, 2021

  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,” in IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS) , 2012, pp. 573–580

  24. [32]

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

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

  25. [33]

    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

Pith tools

Reviewed August 9, 2026 · model on record in the stance chip above.