Pith. sign in

REVIEW 4 major objections 5 minor 50 references

TVG-SLAM: Robust Gaussian Splatting SLAM with Tri-view Geometric Constraints

T0 review · 4 major / 5 minor · reviewed 2026-08-06 · deepseek-v4-flash

Pith's one-line read By adding tri-view geometric constraints to RGB-only Gaussian Splatting SLAM, TVG-SLAM cuts average trajectory error by 69.0% on Cambridge Landmarks while improving rendering quality.

desk verdict Tri-view tracking is a genuine improvement for RGB-only GS SLAM, but the DART mechanism's equation is inverted as printed and has to be fixed before the headline numbers can be taken at face value. read the letter →

arxiv 2506.23207 v1 pith:4BFRGNYR submitted 2025-06-29 cs.CV

classification cs.CV
keywords RGB-onlySLAM3DGaussianSplattingtri-viewmatchingtrifocaltensorcameratrackinguncertainty-guidedinitializationoutdoorvisualdynamicphotometricweighting
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

The paper sets out to show that the main weakness of RGB-only 3D Gaussian Splatting SLAM in outdoor scenes is its dependence on photometric rendering loss for camera tracking, and that this weakness can be removed with a tri-view geometric approach. The method builds dense correspondences across three consecutive frames and uses them to construct two geometric tracking losses, a trifocal tensor-based 2D reprojection loss and a 3D point alignment loss, that complement the photometric loss. With these constraints, an uncertainty-guided Gaussian initialization, and a dynamic attenuation of photometric trust when the map is stale, the system reports 69.0% lower average trajectory error on Cambridge Landmarks than the previous best RGB-only 3DGS baseline, and stronger rendering quality on all three outdoor benchmarks tested. A sympathetic reader would care because these are precisely the conditions, aggressive motion, changing light, and long low-parallax stretches, where current Gaussian Splatting SLAM systems lose track.

What carries the argument

The central object is the tri-view match set $M_{k-1,k,t}$, built by computing dense pairwise matches between the current frame and the latest keyframe, and between the two latest keyframes, then keeping only triplets $(p_{k-1}, p_k, p_t)$ that share the bridge pixel $p_k$. This set carries the system: tracking uses it in the hybrid loss $L_{\mathrm{track}} = \lambda_p L_{\mathrm{photo}} + \lambda_{2D} L_{2D} + \lambda_{3D} L_{3D}$, where $L_{2D}$ is a Huber-weighted algebraic distance between $p_t$ and the epipolar line obtained by transferring $p_{k-1}$ through the trifocal tensor, and $L_{3D}$ is the Procrustes-scaled distance between matching 3D pointmap points. The same triplets supply the per-point tri-view variance $\sigma_g^2$ that the initialization strategy encodes into new Gaussians, with covariance proportional to $\sqrt{\sigma_g^2}$ and opacity decreasing in $\sigma_g^2$, while the dynamic attenuation mechanism sets the photometric weight $\lambda_p$ by a sigmoid of the number of frames since the last keyframe. The machinery's job is to make pose estimation largely independent of rendering quality, so that mapping lag or bad lighting cannot drag the trajectory.

What would settle it

Run TVG-SLAM on an outdoor sequence while corrupting the dense matcher's pointmaps (for example, adding increasing noise to the 3D point positions or replacing the learned matcher with classical feature matching), and record ATE; if trajectory error does not rise toward the photometric-only baseline as pointmap quality drops, then the tri-view geometric constraints are not the cause of the reported improvement. A second check is to evaluate on a sequence with severe lighting changes where pointmap matching is known to fail: the 69.0% improvement should disappear if the paper's explanation is right.

Watch

Extended reading notes

Core claim

The paper claims that consistent tri-view matches, rather than pairwise matches or photometric alignment alone, are the right carrier of geometric signal for RGB-only 3DGS SLAM. By matching the current frame against two recent keyframes and keeping only triplets that agree at the bridge keyframe, the system obtains correspondences with two independent pointmaps, which feed a 2D trifocal constraint (epipolar line transfer) and a 3D alignment loss (Procrustes-scaled pointmap distance). These geometric losses are combined with photometric loss in tracking, while a dynamic weighting mechanism lowers the photometric weight when the rendered map is stale, and a probabilistic initialization encodes tri-view variance into the shape and opacity of new Gaussians. The numerical claim is that this design reduces average absolute trajectory error by 69.0% on Cambridge Landmarks, 65.7% on Small City, and 28.2% on Waymo relative to the prior best RGB-only 3DGS baseline, while improving PSNR across all three benchmarks. The paper interprets these results as evidence that geometric supervision is not a refinement but the load-bearing component for reliable outdoor tracking.

Load-bearing premise

The system assumes the learned dense matcher produces accurate, densely consistent 3D pointmaps and correspondences under exactly the outdoor conditions it targets (low parallax, fast motion, illumination change); if that matcher degrades, both geometric losses lose their signal and tracking falls back on the photometric loss the paper is trying to replace.

Editorial extensions

If this is right

  • On the Waymo driving dataset, average ATE drops by about 28.2% versus the prior best RGB-only 3DGS baseline and by an order of magnitude versus two photometric-only baselines, with average PSNR rising to 25.38 dB.
  • On Small City and Cambridge Landmarks, the reported average ATE reductions are 65.7% and 69.0% respectively, with consistent PSNR, SSIM, and LPIPS improvements over all RGB-only baselines.
  • Ablations attribute the gain to all three components: removing both geometric losses raises ATE on a Waymo scene from 0.870 m to 1.269 m, removing the dynamic attenuation raises it to 1.053 m, and removing the uncertainty-guided initialization raises it to 1.203 m and causes the largest rendering-quality drop (0.72 dB PSNR).
  • Because only matches with adequate parallax are kept for the geometric losses, the system exploits geometric constraints precisely in the low-parallax, straight-line driving cases where photometric methods drift.
  • The dense matching stage currently dominates runtime at about 400 ms per frame, and the authors state that the modular design allows replacing it with lighter matchers for real-time operation.

Reading between the lines

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

  • If the learned matcher's pointmaps are the real source of the geometric signal, then swapping in a cheaper or noisier matcher should require rebalancing $\lambda_{2D}$ and $\lambda_{3D}$; a direct experiment would replace the matcher and measure the ATE and PSNR trade-off.
  • The tri-view variance $\sigma_g^2$ could serve beyond initialization, for example as a criterion to prune or merge redundant Gaussians or to flag dynamic objects, since unstable triplets across three frames are exactly what the matching module already filters out.
  • DART's staleness proxy, the number of frames since the last keyframe, is indirect; measuring actual map drift or rendering error against the current frame could make the photometric attenuation more principled and possibly reduce the 21% ATE penalty the authors report when DART is disabled.
  • The reported gains are on three outdoor datasets with ground-truth poses; testing the same tri-view geometric losses on indoor or handheld AR sequences, where parallax is larger but texture is richer, would show whether the trifocal term remains beneficial or becomes redundant with photometric loss.
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 TVG-SLAM, an RGB-only 3D Gaussian Splatting SLAM system that replaces pure photometric tracking with a hybrid objective combining photometric loss, a trifocal-tensor-based 2D reprojection loss, and a 3D alignment loss built on dense tri-view correspondences obtained from DUST3R. It also introduces two mapping/tracking components: TUGI, which initializes Gaussians with uncertainty derived from multi-view consistency, and DART, which dynamically down-weights the photometric loss when the map is stale. Experiments on Waymo, Small City, and Cambridge Landmarks report large ATE reductions over prior RGB-only 3DGS SLAM baselines, including a 69.0% average ATE reduction on Cambridge Landmarks, with improved rendering metrics. The paper claims that these results demonstrate substantially improved robustness to viewpoint and illumination changes.

Significance. If the results hold, TVG-SLAM provides strong evidence that dense multi-view geometric constraints, rather than photometric rendering alone, can stabilize RGB-only 3DGS SLAM in challenging outdoor conditions. The use of tri-view correspondences to directly supervise both 2D and 3D pose optimization is a sensible and potentially impactful direction, and the paper presents ablations showing that each proposed component contributes to tracking and rendering quality. However, the manuscript currently contains an internal contradiction in the DART equation, missing hyperparameters for TUGI, and insufficient specification of the geometric filtering procedure. These issues prevent the reader from verifying or reproducing the central claims as written. The paper also does not release code or provide full implementation details, which further limits reproducibility.

major comments (4)
  1. [III-D, Eq. (7)] The DART mechanism is described as a decreasing sigmoid-like function of map staleness ΔNf, but Eq. (7) as written defines λp = wmin + (wmax − wmin)σ(k(ΔNf − Nm)) with σ as a standard increasing sigmoid. This makes λp increase from near wmin to wmax as staleness grows, which is exactly opposite of the described attenuation. The correct form would be σ(−k(ΔNf − Nm)) or an explicit 1 − σ(...) term. This is not a cosmetic issue: the paper credits DART with preventing tracking failure during aggressive motion (Sec. IV-A) and reports a 21% ATE increase when DART is removed (Table IV). The manuscript must correct Eq. (7) and clarify the sigmoid definition (the text states σ(x) = 1/(1 + ex), which is ambiguous; presumably e^{-x} is intended). Without this correction, the DART contribution as specified does not behave as claimed.
  2. [III-E.2 (TUGI)] The TUGI opacity formula is not a valid probability definition and is missing parameter values. The paper states αnew = sigmoid−1(a(1−k√σ2_g)), but the inverse sigmoid (logit) maps to (−∞, ∞), not [0, 1], and would be an unusual way to define an opacity. This appears to be a typo for sigmoid(a(1−k√σ2_g)) or similar. Moreover, the hyperparameters a and k for the opacity modulation, as well as the proportionality constant in Snew ∝ √σ2_g, are never specified in the Implementation Details. Since the ablation in Table IV shows that removing TUGI causes the largest degradation in both ATE (+38.3%) and PSNR (−0.72 dB), the missing parameter values directly affect reproducibility and must be provided.
  3. [III-C and III-D] The manuscript repeatedly refers to using only matches that satisfy 'favorable geometric conditions (e.g., sufficient parallax)' and mentions a 'geometric filter threshold' in the free parameters, but no concrete criterion, threshold value, or filtering procedure is given. The entire tracking loss L2D and L3D is computed on the filtered tri-view correspondence set, so without this information the method cannot be reimplemented or fairly compared. Please specify the exact filter, including any parallax angle thresholds, confidence thresholds on DUST3R matches, or outlier rejection steps.
  4. [IV-B, Table IV] The ablation study for DART and the geometric constraints is performed on only one Waymo scene (153495), and the additional DART analysis in Fig. 5 uses a different scene (158686). Given that the paper's headline claim is a 69% ATE reduction on Cambridge Landmarks and DART is specifically motivated by aggressive hand-held motion, the current evidence is too thin to support the generality of DART's contribution. At minimum, the authors should add a DART ablation on the Cambridge Landmarks or Small City datasets, or explicitly justify why a single scene is sufficient.
minor comments (5)
  1. [Eq. (8)] The uncertainty variance formula would be clearer as σ2_g = (1/N) Σ_i ||P_i − P̄||² to explicitly show the Euclidean norm; as written, it may be misread as a plain squared difference.
  2. [Reference [50]] The Cambridge Landmarks dataset is cited to a Hyperpose paper (arXiv:2303.02610), but that is not the original dataset paper. The proper reference is likely Kendall et al., 'PoseNet: A Convolutional Network for Real-Time 6-DOF Camera Relocalization' (ICCV 2015), which introduced Cambridge Landmarks. Please correct this citation.
  3. [Table III] For the Cambridge Landmarks seq 4, MonoGS reports NAN for PSNR, SSIM, and LPIPS but an ATE value of 6.701. The table caption and average row should clarify how these missing values are treated in the average (e.g., excluded from the mean).
  4. [Sec. I and Table V] The paper emphasizes real-time rendering as a motivation for 3DGS, but the proposed system's dense matching stage alone takes 400 ms per frame, and the full pipeline is not real-time. The limitation is acknowledged in Sec. IV-B, but it would be more transparent to state explicitly in the abstract or introduction that the current system is not real-time and the matching step is the bottleneck.
  5. [Sec. III-D, Eq. (5)] The algebraic residual in Eq. (5) is squared, but the overall loss is wrapped in a Huber loss ρ(·). If ρ is applied to the squared residual sum, the quadratic behavior may be redundant; please clarify the exact form of ρ and where it is applied.

Circularity Check

0 steps flagged · score 1.0 of 10

No significant circularity: the tracking and mapping losses are supervised by external DUST3R pointmaps and standard trifocal geometry, and the headline ATE and rendering metrics are independent external benchmarks; the self-citations are non-load-bearing related-work mentions.

full rationale

TVG-SLAM's derivation is not circular in any of the enumerated senses. The tri-view matching module uses DUST3R's pretrained pairwise pointmaps as an external geometric source; the tracking losses L2D and L3D (Eqs. 5 and 6) minimize algebraic and Procrustes residuals against those pointmaps and standard trifocal geometry, while Lphoto (Eq. 3) is a photometric residual against the rendered map. None of these objectives defines the reported quantities (ATE RMSE, PSNR, SSIM, LPIPS), which are computed against ground-truth trajectories and reference images on Waymo, Small City, and Cambridge Landmarks. The implementation details disclose hyperparameters and borrowed components (relative-scale accumulation from OpenGS-SLAM) rather than fitting a parameter that is renamed as the headline result. The self-citations [10] and [34] appear only in related-work surveys of pose optimization and are not used to justify the tri-view constraints, DART, or TUGI. One internal consistency issue is worth flagging separately from circularity: Eq. 7 writes lambda_p = w_min + (w_max - w_min) * sigma(k(Delta_N_f - N_m)) with sigma increasing, so as printed lambda_p rises with map staleness, opposite to the stated decreasing attenuation; this is a correctness/specification error in the DART component, not a circular definition, and it does not change the verdict that the evaluated metrics are independent of the method's own equations.

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

The central claim rests on standard multi-view geometry, on DUST3R's pretrained predictions, and on several hand-set hyperparameters that are not fully specified. No new physical or model entities are introduced.

free parameters (5)
  • lambda_2D and lambda_3D = 0.01 each
    Weights in Eq. 2 set "empirically" in Section IV; no validation protocol is described.
  • DART parameters (w_min, w_max, N_m, k) = 0.1, 1.0, 5, 0.8
    Chosen by hand in the Implementation Details; no sensitivity analysis is provided.
  • TUGI opacity parameters a and k = not specified
    Introduced in Section III-E.2 but no numeric values are given in the paper.
  • TUGI covariance scaling constant = S_new proportional to sqrt(sigma_g^2)
    The proportionality constant is unspecified.
  • Geometric filter threshold = not specified
    Section III-D says matches must satisfy "favorable geometric conditions (e.g., sufficient parallax)" without giving the threshold or rule.
assumptions (5)
  • standard math Trifocal tensor geometry and epipolar line transfer (Hartley-Zisserman [45])
    Used in Eq. 4-5 for the 2D constraint; standard multi-view geometry.
  • domain assumption DUST3R provides accurate dense pointmaps and correspondences in outdoor scenes
    Section III-C and III-D rely on DUST3R for all geometric constraints; no failure analysis for the target illumination and viewpoint conditions is provided.
  • domain assumption Relative scale accumulation from OpenGS-SLAM [26] keeps tri-view pointmaps in a consistent metric frame
    Invoked in Implementation Details and Eq. 6; if scale drifts, the 3D alignment loss is invalid.
  • domain assumption The number of frames since the last keyframe (Delta N_f) is a reliable proxy for map staleness
    DART in Eq. 7 models photometric trust as a function of Delta N_f; the paper provides no direct evidence linking this proxy to actual map error.
  • domain assumption Triplet construction by sharing pixel p_k in the intermediate keyframe yields geometrically consistent triplets
    Section III-C step 2 searches for a match sharing p_k; subpixel matching errors could break consistency.

how reviews work

0 comments
Cite this review

Pith. "Pith review of TVG-SLAM: Robust Gaussian Splatting SLAM with Tri-view Geometric Constraints." pith.science (2026). https://pith.science/paper/4BFRGNYR

@misc{pith2026250623207,
  author       = {Pith},
  title        = {Pith review of: TVG-SLAM: Robust Gaussian Splatting SLAM with Tri-view Geometric Constraints},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/4BFRGNYR}},
  note         = {Machine review of arXiv:2506.23207}
}
read the original abstract

Recent advances in 3D Gaussian Splatting (3DGS) have enabled RGB-only SLAM systems to achieve high-fidelity scene representation. However, the heavy reliance of existing systems on photometric rendering loss for camera tracking undermines their robustness, especially in unbounded outdoor environments with severe viewpoint and illumination changes. To address these challenges, we propose TVG-SLAM, a robust RGB-only 3DGS SLAM system that leverages a novel tri-view geometry paradigm to ensure consistent tracking and high-quality mapping. We introduce a dense tri-view matching module that aggregates reliable pairwise correspondences into consistent tri-view matches, forming robust geometric constraints across frames. For tracking, we propose Hybrid Geometric Constraints, which leverage tri-view matches to construct complementary geometric cues alongside photometric loss, ensuring accurate and stable pose estimation even under drastic viewpoint shifts and lighting variations. For mapping, we propose a new probabilistic initialization strategy that encodes geometric uncertainty from tri-view correspondences into newly initialized Gaussians. Additionally, we design a Dynamic Attenuation of Rendering Trust mechanism to mitigate tracking drift caused by mapping latency. Experiments on multiple public outdoor datasets show that our TVG-SLAM outperforms prior RGB-only 3DGS-based SLAM systems. Notably, in the most challenging dataset, our method improves tracking robustness, reducing the average Absolute Trajectory Error (ATE) by 69.0\% while achieving state-of-the-art rendering quality. The implementation of our method will be released as open-source.

Figures

Figures reproduced from arXiv: 2506.23207 by the authors.

Figure 1
Figure 1. Our system integrates tri-view geometric constraints to achieve [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. The TVG-SLAM pipeline. Our system processes incremental RGB images by first building robust tri-view matches. In tracking, [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗
Figure 3
Figure 3. Rendering quality comparison on the Waymo, Small City, and Cambridge Landmarks datasets in unbounded outdoor scenes. Compared [PITH_FULL_IMAGE:figures/full_fig_p005_3.png] view at source ↗
Figures from the paper (2 more)
Figure 4
Figure 4. Figure 4: Trajectory comparison on challenging outdoor sequences. [PITH_FULL_IMAGE:figures/full_fig_p007_4.png]
Figure 5
Figure 5. Figure 5: Ablation in DART on the Waymo dataset (scene: 158686). [PITH_FULL_IMAGE:figures/full_fig_p007_5.png]

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

50 extracted references · 41 canonical work pages

  1. [1]

    Past, present, and future of simultaneous localization and mapping: Toward the robust-perceptual age,

    C. Cadena, L. Carlone, H. Carrillo, Y . Latif, D. Scaramuzza, J. Neira, I. Reid, and J. J. Leonard, “Past, present, and future of simultaneous localization and mapping: Toward the robust-perceptual age,” IEEE Transactions on robotics, vol. 32, no. 6, pp. 1309–1332, 2016

  2. [2]

    Dtam: Dense tracking and mapping in real-time,

    R. A. Newcombe, S. J. Lovegrove, and A. J. Davison, “Dtam: Dense tracking and mapping in real-time,” in Proc. of the IEEE/CVF Intl. Conf. on Computer Vision (ICCV) , 2011

  3. [3]

    Slam++: Simultaneous localisation and mapping at the level of objects,

    R. F. Salas-Moreno, R. A. Newcombe, H. Strasdat, P. H. Kelly, and A. J. Davison, “Slam++: Simultaneous localisation and mapping at the level of objects,” in Proc. of the IEEE/CVF Conf. on Computer Vision and Pattern Recognition (CVPR), 2013

  4. [4]

    Orb-slam2: An open-source slam system for monocular, stereo, and rgb-d cameras,

    R. Mur-Artal and J. D. Tard ´os, “Orb-slam2: An open-source slam system for monocular, stereo, and rgb-d cameras,” IEEE Trans. on Robotics (TRO), vol. 33, no. 5, pp. 1255–1262, 2017

  5. [5]

    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

  6. [6]

    Elasticfusion: Dense slam without a pose graph

    T. Whelan, S. Leutenegger, R. F. Salas-Moreno, B. Glocker, and A. J. Davison, “Elasticfusion: Dense slam without a pose graph.” in Robotics: Science and Systems , vol. 11, 2015, p. 3

  7. [7]

    Kinect- fusion: Real-time dense surface mapping and tracking,

    R. A. Newcombe, S. Izadi, O. Hilliges, D. Molyneaux, D. Kim, A. J. Davison, P. Kohi, J. Shotton, S. Hodges, and A. Fitzgibbon, “Kinect- fusion: Real-time dense surface mapping and tracking,” in Proc. of the Intl. Symposium on Mixed and Augmented Reality (ISMAR) , 2011

  8. [8]

    Codeslam—learning a compact, optimisable representation for dense visual slam,

    M. Bloesch, J. Czarnowski, R. Clark, S. Leutenegger, and A. J. Davison, “Codeslam—learning a compact, optimisable representation for dense visual slam,” in Proc. of the IEEE/CVF Conf. on Computer Vision and Pattern Recognition (CVPR), 2018

Show all 50 references
  1. [9]

    Bad slam: Bundle adjusted direct rgb-d slam,

    T. Schops, T. Sattler, and M. Pollefeys, “Bad slam: Bundle adjusted direct rgb-d slam,” in Proc. of the IEEE/CVF Conf. on Computer Vision and Pattern Recognition (CVPR) , 2019

  2. [10]

    Pipo-slam: Lightweight visual- inertial slam with preintegration merging theory and pose-only de- scriptions of multiple view geometry,

    Y . Ge, L. Zhang, Y . Wu, and D. Hu, “Pipo-slam: Lightweight visual- inertial slam with preintegration merging theory and pose-only de- scriptions of multiple view geometry,” IEEE Transactions on Robotics , vol. 40, pp. 2046–2059, 2024

  3. [11]

    imap: Implicit mapping and positioning in real-time,

    E. Sucar, S. Liu, J. Ortiz, and A. J. Davison, “imap: Implicit mapping and positioning in real-time,” in Proc. of the IEEE/CVF Intl. Conf. on Computer Vision (ICCV) , 2021

  4. [12]

    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 Proc. of the IEEE/CVF Conf. on Computer Vision and Pattern Recognition (CVPR), 2022

  5. [13]

    Neus: learning neural implicit surfaces by volume rendering for multi- view reconstruction,

    P. Wang, L. Liu, Y . Liu, C. Theobalt, T. Komura, and W. Wang, “Neus: learning neural implicit surfaces by volume rendering for multi- view reconstruction,” in Proc. of the Conference on Neural Information Processing Systems (NeurIPS) , 2021

  6. [14]

    Orbeez-slam: A real-time monocular visual slam with orb features and nerf-realized mapping,

    C.-M. Chung, Y .-C. Tseng, Y .-C. Hsu, X.-Q. Shi, Y .-H. Hua, J.-F. Yeh, W.-C. Chen, Y .-T. Chen, and W. H. Hsu, “Orbeez-slam: A real-time monocular visual slam with orb features and nerf-realized mapping,” in Proc. of the IEEE Intl. Conf. on Robotics & Automation (ICRA) , 2023

  7. [15]

    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 Proc. of the IEEE/CVF Conf. on Computer Vision and Pattern Recognition (CVPR) , 2023

  8. [16]

    Go-slam: Global optimization for consistent 3d instant reconstruction,

    Y . Zhang, F. Tosi, S. Mattoccia, and M. Poggi, “Go-slam: Global optimization for consistent 3d instant reconstruction,” in Proc. of the IEEE/CVF Intl. Conf. on Computer Vision (ICCV) , 2023

  9. [17]

    End-to-end rgb-d slam with multi-mlps dense neural implicit representations,

    M. Li, J. He, Y . Wang, and H. Wang, “End-to-end rgb-d slam with multi-mlps dense neural implicit representations,” IEEE Robotics and Automation Letters (RA-L) , 2023

  10. [18]

    Neuv-slam: Fast neural mul- tiresolution voxel optimization for rgbd dense slam,

    W. Guo, B. Wang, and L. Chen, “Neuv-slam: Fast neural mul- tiresolution voxel optimization for rgbd dense slam,” arXiv preprint arXiv:2402.02020, 2024

  11. [19]

    Nicer-slam: Neural implicit scene encoding for rgb slam,

    Z. Zhu, S. Peng, V . Larsson, Z. Cui, M. R. Oswald, A. Geiger, and M. Pollefeys, “Nicer-slam: Neural implicit scene encoding for rgb slam,” in 2024 International Conference on 3D Vision (3DV) . IEEE, 2024, pp. 42–52

  12. [20]

    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 Trans. Graph. , vol. 42, no. 4, 2023

  13. [21]

    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 Proc. of the IEEE/CVF Conf. on Computer Vision and Pattern Recognition (CVPR) , 2024

  14. [22]

    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

  15. [23]

    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 arXiv preprint arXiv:2312.02126 , 2023

  16. [24]

    Gaussian splatting slam,

    H. Matsuki, R. Murai, P. H. Kelly, and A. J. Davison, “Gaussian splatting slam,” in Proc. of the IEEE/CVF Conf. on Computer Vision and Pattern Recognition (CVPR), 2024

  17. [25]

    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 arXiv preprint arXiv:2312.11284 , 2023

  18. [26]

    Opengs-slam: A cross-platform open-source system for gaussian splatting-based slam,

    Z. Wang, Y . Lan, Y . Chen, Y . Liu, Y . Liu, Y . Chen, and G. Huang, “Opengs-slam: A cross-platform open-source system for gaussian splatting-based slam,” arXiv preprint arXiv:2407.06555 , 2024

  19. [27]

    Direct sparse odometry,

    J. Engel, V . Koltun, and D. Cremers, “Direct sparse odometry,” IEEE TPAMI, 2018

  20. [28]

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

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

  21. [29]

    Nerf– : Neural radiance fields without known camera parameters,

    Z. Wang, S. Wu, W. Xie, M. Chen, and V . A. Prisacariu, “Nerf– : Neural radiance fields without known camera parameters,” arXiv preprint arXiv:2102.07064, 2021

  22. [30]

    Barf: Bundle-adjusting neural radiance fields,

    C.-H. Lin, W.-C. Ma, A. Torralba, and S. Lucey, “Barf: Bundle-adjusting neural radiance fields,” in Proceedings of the IEEE/CVF international conference on computer vision , 2021, pp. 5742–5751

  23. [31]

    Self- calibrating neural radiance fields,

    Y . Jeong, S. Ahn, C. Choy, A. Anandkumar, M. Cho, and J. Park, “Self- calibrating neural radiance fields,” in Proceedings of the IEEE/CVF International Conference on Computer Vision , 2021, pp. 5846–5854

  24. [32]

    Local-to-global registration for bundle-adjusting neural radiance fields,

    Y . Chen, X. Chen, X. Wang, Q. Zhang, Y . Guo, Y . Shan, and F. Wang, “Local-to-global registration for bundle-adjusting neural radiance fields,” in Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2023, pp. 8264–8273

  25. [33]

    Nope-nerf: Optimising neural radiance field with no pose prior,

    W. Bian, Z. Wang, K. Li, J.-W. Bian, and V . A. Prisacariu, “Nope-nerf: Optimising neural radiance field with no pose prior,” in Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2023, pp. 4160–4169

  26. [34]

    Td-nerf: Novel truncated depth prior for joint camera pose and neural radiance field optimization,

    Z. Tan, Z. Zhou, Y . Ge, Z. Wang, X. Chen, and D. Hu, “Td-nerf: Novel truncated depth prior for joint camera pose and neural radiance field optimization,” in 2024 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS) . IEEE, 2024, pp. 372–379

  27. [35]

    RA-NeRF: Towards high-fidelity and robust NeRF-based SLAM with monocular camera,

    Z. Yuan, X. Chen, J. Wang, S. Li, and T. Zhang, “RA-NeRF: Towards high-fidelity and robust NeRF-based SLAM with monocular camera,” in 2024 IEEE International Conference on Robotics and Automation (ICRA). IEEE, 2024, pp. 9172–9179

  28. [36]

    Ct-nerf: Incremental optimizing neural radiance field and poses with complex trajectory,

    Y . Ran, Y . Li, Q. Ye, Y . Huo, Z. Bai, J. Sun, and J. Chen, “Ct-nerf: Incremental optimizing neural radiance field and poses with complex trajectory,” arXiv preprint arXiv:2404.13896 , 2024

  29. [37]

    Cf- nerf: Camera parameter free neural radiance fields with incremental learning,

    Q. Yan, Q. Wang, K. Zhao, J. Chen, B. Li, X. Chu, and F. Deng, “Cf- nerf: Camera parameter free neural radiance fields with incremental learning,” in Proceedings of the AAAI Conference on Artificial Intel- ligence, vol. 38, no. 6, 2024, pp. 6440–6448

  30. [38]

    Colmap- free 3d gaussian splatting,

    Y . Fu, S. Liu, A. Kulkarni, J. Kautz, A. A. Efros, and X. Wang, “Colmap- free 3d gaussian splatting,” in Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , 2024, pp. 20 796–20 805

  31. [39]

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

    X. Yang, H. Li, H. Zhai, Y . Ming, Y . Liu, and G. Zhang, “V ox- fusion: Dense tracking and mapping with voxel-based neural implicit representation,” in Proc. of the Intl. Symposium on Mixed and Augmented Reality (ISMAR), 2022

  32. [40]

    Loopy-slam: Dense neural slam with loop closures,

    L. Liso, E. Sandstr ¨om, V . Yugay, L. Van Gool, and M. R. Oswald, “Loopy-slam: Dense neural slam with loop closures,” in Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2024, pp. 20 363–20 373

  33. [41]

    Splat-slam: Globally optimized rgb-only slam with 3d gaussians,

    E. Sandstr ¨om, G. Zhang, K. Tateno, M. Oechsle, M. Niemeyer, Y . Zhang, M. Patel, L. Van Gool, M. Oswald, and F. Tombari, “Splat-slam: Globally optimized rgb-only slam with 3d gaussians,” in Proceedings of the Computer Vision and Pattern Recognition Conference , 2025, pp. 1680–1691

  34. [42]

    Mgs-slam: Monocular sparse tracking and gaussian mapping with depth smooth regularization,

    P. Zhu, Y . Zhuang, B. Chen, L. Li, C. Wu, and Z. Liu, “Mgs-slam: Monocular sparse tracking and gaussian mapping with depth smooth regularization,” IEEE Robotics and Automation Letters , 2024

  35. [43]

    Motiongs: Compact gaussian splatting slam by motion filter,

    X. Guo, W. Zhang, R. Liu, P. Han, and H. Chen, “Motiongs: Compact gaussian splatting slam by motion filter,” in 2024 7th International Conference on Robotics, Control and Automation Engineering (RCAE) . IEEE, 2024, pp. 685–692

  36. [44]

    Dust3r: Geometric 3d vision made easy,

    S. Wang, V . Leroy, Y . Cabon, B. Chidlovskii, and J. Revaud, “Dust3r: Geometric 3d vision made easy,” in Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , 2024, pp. 20 697–20 709

  37. [45]

    R. I. Hartley and A. Zisserman, Multiple View Geometry in Computer Vision, 2nd ed. Cambridge University Press, ISBN: 0521540518, 2004

  38. [46]

    Incremental light bundle adjustment,

    V . Indelman, S. Williams, J. Gallier, and F. Dellaert, “Incremental light bundle adjustment,” in IEEE International Conference on Robotics and Automation (ICRA). IEEE, 2012, pp. 5364–5371

  39. [47]

    Least-squares estimation of transformation parameters between two point patterns,

    S. Umeyama, “Least-squares estimation of transformation parameters between two point patterns,” IEEE Transactions on Pattern Analysis & Machine Intelligence, vol. 13, no. 04, pp. 376–380, 1991

  40. [48]

    Scalability in perception for autonomous driving: Waymo open dataset,

    P. Sun, H. Kretzschmar, X. Dotiwalla, A. Chouard, V . Patnaik, P. Tsui, J. Guo, Y . Zhou, Y . Chai, B. Caine, et al. , “Scalability in perception for autonomous driving: Waymo open dataset,” in Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, ...

  41. [49]

    A hierarchical 3d gaussian representation for real-time rendering of very large datasets,

    B. Kerbl, A. Meuleman, G. Kopanas, M. Wimmer, A. Lanvin, and G. Drettakis, “A hierarchical 3d gaussian representation for real-time rendering of very large datasets,”ACM Transactions on Graphics (TOG), vol. 43, no. 4, pp. 1–15, 2024

  42. [50]

    Hyperpose: Camera pose localization using attention hypernetworks,

    R. Ferens and Y . Keller, “Hyperpose: Camera pose localization using attention hypernetworks,” arXiv preprint arXiv:2303.02610 , 2023

Pith tools

Reviewed August 6, 2026 · model on record in the stance chip above.