Pith. sign in

REVIEW 3 major objections 6 minor 1 cited by

Outdoor Monocular SLAM with Global Scale-Consistent 3D Gaussian Pointmaps

T0 review · 3 major / 6 minor · reviewed 2026-08-06 · deepseek-v4-flash

Pith's one-line read S3PO-GS estimates camera poses from 2D-3D correspondences between the current frame and the map's own rendered pointmaps, inheriting the 3D Gaussian scene's scale and avoiding cumulative outdoor scale drift.

desk verdict Useful 3DGS SLAM system with a clever pose-anchoring idea, but the scale-drift claims need better evidence before they can be believed. read the letter →

arxiv 2507.03737 v2 pith:IA4BIUCD submitted 2025-07-04 cs.CV

classification cs.CV
keywords 3DGaussianSplattingmonocularSLAMscaledriftpointmapnovelviewsynthesisoutdoorscenescameraposetrackinggeometricpriors
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

This paper claims that the persistent scale-drift problem in monocular outdoor 3D Gaussian Splatting SLAM can be removed by anchoring pose estimation in the map's own rendered pointmaps rather than in the scale of an external tracking model. The system, S3PO-GS, renders a depth-based pointmap from the adjacent keyframe through the 3DGS rasterizer, uses a pre-trained pointmap model only to find pixel correspondences across frames, and then solves a perspective-n-point problem against the map's own 3D coordinates. Because those coordinates carry the Gaussian scene's scale, the estimated pose is scale-consistent by construction, and the pre-trained model never feeds scale into the pose. A patch-based dynamic mapping module then aligns the pre-trained pointmap's scale locally to the scene and replaces incorrect points before inserting Gaussians, giving the geometry supervision that RGB-only outdoor scenes otherwise lack. On three outdoor benchmarks, the paper reports state-of-the-art novel view synthesis and tracking that is substantially more accurate than prior 3DGS SLAM baselines, with accurate poses already at five optimization iterations.

What carries the argument

The load-bearing object is the 3DGS-rendered pointmap: a per-pixel map of 3D coordinates obtained by $\alpha$-blending the Gaussian depth along each ray and unprojecting with the inverse camera intrinsics. Its role is to make the Gaussian map itself the geometric reference: correspondences from the current frame are matched to this pointmap, so the PnP solution inherits the map's scale and no external scale factor needs to be maintained. The second mechanism is the patch-based scale alignment: the rendered pointmap $X^r$ and the pre-trained pointmap $X^p$ are cut into $P\times P$ patches, patches with similar mean and standard deviation are normalized, points agreeing after normalization become correct points, and their mean ratio gives a scale factor that aligns $X^p$ to the scene; unreliable points in $X^r$ are then replaced with aligned values before new Gaussians are inserted.

What would settle it

Mask out a large central portion of the rendered pointmap in the adjacent keyframe on a driving sequence and measure the ATE; if tracking error does not rise substantially when the pointmap anchor is missing, the claimed reliance on pointmap-anchored 2D-3D correspondences would be falsified.

Watch

Extended reading notes

Core claim

The central claim is that a 3DGS SLAM system can make its own map the source of scale for camera tracking. At each new frame, S3PO-GS rasterizes the adjacent keyframe into a depth map, unprojects it into a pointmap $X^r$, and uses the pre-trained pointmap model to establish pixel correspondences between the keyframe and the current frame. The scale of $X^r$ comes from the Gaussian map itself, so the 2D-3D correspondences fed to RANSAC and PnP are already in the scene's own metric scale, and the relative pose $T_{\rm rel}$ solved from them preserves that scale; the pre-trained model only supplies correspondence, not pose. Tracking is then refined through a differentiable photometric loss in the 3DGS rasterizer. For mapping, the pre-trained pointmap $X^p$ for a keyframe is aligned to the scene by comparing small patches of $X^r$ and $X^p$ through normalized statistics, selecting reliable points, computing a scaling factor, and iterating until stable; misaligned points in the rendered pointmap are replaced before new Gaussians are inserted, and a pointmap geometry loss supervises map optimization. The reported result is that this yields state-of-the-art novel view synthesis on all three outdoor datasets, tracking errors reduced by 67.5% and 77.3% against the leading outdoor 3DGS SLAM baseline on two of them, and tracking that converges in about 10% of the iterations.

Load-bearing premise

The rendered pointmap from the adjacent keyframe must be accurate and complete enough to give reliable 2D-3D correspondences; if the Gaussian map has holes, dynamic-object artifacts, or wrong-scale regions in view, the PnP pose inherits those errors, so the claim of avoiding cumulative scale drift depends on the map already being correct at each step.

Editorial extensions

If this is right

  • Accurate poses become available with roughly one tenth of the optimization iterations needed by tracking that relies on differentiable rendering alone, since the pointmap anchor gives a strong geometric initialization.
  • Scale stays tied to the 3DGS map rather than to an external tracking model, so cumulative scale-alignment errors in long outdoor sequences are structurally avoided rather than patched.
  • Novel view synthesis improves on all three outdoor benchmarks, with the reported gains indicating that geometric pointmap supervision helps reconstruct complex depth boundaries.
  • Removing any of the three mapping components (scale alignment, point replacement, geometry loss) measurably hurts tracking or reconstruction, so the claimed improvement is carried jointly by the tracking anchor and the patch-based mapping.

Reading between the lines

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

  • An implication the authors leave implicit is that the tracking and mapping loops are coupled: the pointmap anchor is only as reliable as the Gaussian map it renders, so failures such as holes, dynamic objects, or wrong-scale regions in the map would be inherited by the pose estimate; this makes the stated limitation on dynamic scenes a structural one.
  • Because the pre-trained pointmap model acts only as a correspondence bridge, the design could in principle accept any future dense correspondence model that outputs per-pixel 3D coordinates, without changing the scale-consistency argument.
  • The patch-based scale-alignment routine is a local calibration primitive that could be lifted to other monocular mapping systems: wherever two dense pointmap sources disagree in scale, matching patches by normalized statistics before computing a global scale factor may be more robust than whole-image alignment.
  • A natural testable extension would be longer sequences with loop closure or global bundle adjustment; the paper notes these are absent, and adding them could show whether the scale-consistency benefit persists when the scene is revisited.
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

3 major / 6 minor

Summary. The paper proposes S3PO-GS, a monocular RGB-only 3D Gaussian Splatting SLAM pipeline for outdoor scenes. Tracking is performed by rendering a pointmap from the current 3DGS map at the adjacent keyframe, matching the new frame to that keyframe with MASt3R, and solving PnP against the rendered pointmap; the pose is then refined with photometric optimization. Mapping uses a patch-based scale-alignment procedure that rescales MASt3R pointmaps to the current scene scale, replaces inconsistent rendered points, and supervises map optimization with a geometric loss. Experiments on Waymo, KITTI, and DL3DV report state-of-the-art novel view synthesis and tracking accuracy relative to selected 3DGS-SLAM baselines, with ablations showing the contribution of each module.

Significance. If the empirical claims hold, the paper would make a worthwhile contribution: anchor-based PnP tracking from a differentiable 3DGS pointmap is a plausible way to reduce pose-optimization iterations by an order of magnitude, and the patch-based alignment plus point replacement appears to improve reconstruction quality. The writing is clear, the method is described in enough detail to reproduce the main pipeline, and the supplement provides algorithms, hyperparameters, and an honest list of limitations (dynamic objects, no loop closure, no global BA). However, the central claim of avoiding global scale drift is not supported by the current evidence: the scale alignment routine closes the loop on the map's own scale, so no independent scale reference remains after initialization. The ATE alignment protocol is unspecified, results are single-run without error bars, and code/data are not released. These issues must be addressed before the scale-consistency claim can be accepted.

major comments (3)
  1. [3.3.1, 3.3.3, Eqs. (13)-(15)] The scale-alignment procedure rescales the pretrained pointmap X^p to match the rendered pointmap X^r using sigma' = mean(X^r[CP])/mean(X^p[CP]) (Eq. 13), and only then is the aligned pointmap used in L_geo (Eq. 15). Consequently, a global scale error in X^r is invisible to L_geo after alignment, and L_pho is invariant to a joint rescaling of the map and camera poses. The MASt3R prior thus anchors scale only at initialization; after the first frames it is conformed to the map rather than acting as an external reference. The paper's claims of 'global scale-consistent' and 'avoids cumulative scale drift' are therefore not established by the presented equations. To support the claim, the authors should report ATE after SE(3) alignment and after Sim(3) alignment separately, and plot a running scale-ratio metric (e.g., estimated trajectory length divided by ground-truth length) over sequence time. If Sim(3) ATE is substantially lower than SE(3) ATE, the apparent accuracy is at least partly due to alignment absorbing scale drift.
  2. [4.1, Tables 1-5] The ATE RMSE values in Table 1 are presented without specifying whether alignment was performed with SE(3) or Sim(3). This is a critical omission for a monocular pipeline whose central claim is scale consistency, because Sim(3) alignment explicitly absorbs a global scale factor. Moreover, all tables report single-run numbers without error bars or multiple-seed statistics; for DL3DV the evaluation uses only three sequences, so the reported 77.3% tracking improvement over OpenGS-SLAM may not be statistically meaningful. The authors should provide the alignment protocol, per-sequence results, and variance across runs (or at least across dataset splits). Releasing code and data would also allow independent verification of these numbers.
  3. [3.2.1, Eqs. (6)-(9)] The pose estimation chain X^r_ak -> I_ak -> I_n (Eq. 8) assumes that the rendered pointmap X^r_ak and the MASt3R pointmap X^p_ak are pixel-wise spatially consistent at the matched keyframe pixels. Since X^p is later scaled to X^r (Eq. 13), any error in X^r -- holes, dynamic-object artifacts, or poorly optimized Gaussians -- is inherited by the PnP pose. The paper does not analyze how often the rendered pointmap is the limiting factor, nor does it quantify the effect of incomplete or erroneous rendered depth on downstream tracking. An ablation that replaces the rendered pointmap in PnP with the scaled MASt3R pointmap (and vice versa) would help clarify whether the claimed robustness is due to the 3DGS anchor or to the quality of the MASt3R matching itself.
minor comments (6)
  1. [5, Table 1] The conclusion states that the method achieves 'centimeter-level tracking accuracy on complex datasets like Waymo,' but Table 1 reports ATE RMSE of 0.622 m on Waymo and 1.048 m on KITTI, which are tens of centimeters. Please correct the wording or specify a separate metric that supports the centimeter-level claim.
  2. [Figure 1] The figure caption says the results are 'on KITTI,' but the reported ATE values (e.g., 0.473) do not match the KITTI average in Table 1 (1.048). Please clarify whether Figure 1 shows a single representative sequence or a different aggregation than Table 1.
  3. [Supplementary Section 8] The cross-references to 'Tab. 8' in Sections 8.1 and 8.2 appear to point to the running-time table (Table 7), while the patch-size table is actually Table 8. Please correct the table numbering and references.
  4. [Supplementary Section 8.1] The comparison table lists 'MASt3R-SLAM [19]', but reference [19] is the MASt3R paper by Leroy et al., not MASt3R-SLAM; the MASt3R-SLAM citation should be to Murai et al. (reference [27] in the main text).
  5. [3.1, Eq. (3)] The text says z_k is 'the distance from point ui to the camera center,' but the variable ui is not defined; it should refer to the 3D point along the ray corresponding to pixel (i,j). Please fix the notation.
  6. [Supplementary Table 6 and Section 7.4] The hyperparameter table does not list epsilon_m, alpha, or lambda_iso, even though these are used in Eqs. (14) and (17) and values are given in the text. Please add them to Table 6 for completeness.

Circularity Check

2 steps flagged · score 6.0 of 10

Eqs. 13–15 align the pretrained pointmap to the rendered map and then supervise the map with that aligned prior, so global scale becomes a self-consistency constraint rather than an externally anchored prediction; the claimed scale-consistent tracking inherits the map's scale by construction.

  1. fitted input called prediction [Sec. 3.3.1 (Eq. 13) and Sec. 3.3.3 (Eq. 15)]
    "Points satisfying |X r N (x) − X p N (x)| < ϵr are selected as “correct points” set CP , and scale factor is calculated: σ′ = µ(X r[CP ])/µ(X p[CP ]), (13) ... eventually outputting the aligned ˆX p = σ × X p for pointmap replacement and supervision. ... To improve the scene geometry, we introduce a pointmap-based geometry loss: Lgeo = X r − ˆX p 1. (15)"

    The scale factor σ′ is computed from the rendered pointmap X^r itself (mean over selected correct points). The supervised target is then X̂^p = σ′ X^p, i.e., the pretrained pointmap rescaled to X^r's current global scale. Substituting Eq. 13 into Eq. 15 gives Lgeo = ||X^r − (μ(X^r[CP])/μ(X^p[CP])) X^p||_1; multiplying X^r by a global scale s multiplies μ(X^r[CP]) and hence σ′ by the same s, so Lgeo is invariant to a global rescaling of the map. The geometry loss therefore cannot detect or correct a global scale error in the Gaussian map: the 'scale alignment' is an alignment of the prior to the map, and the claimed avoidance of cumulative scale drift is a self-consistency property of the gauge, not an independent scale prediction.

  2. self definitional [Sec. 3.2.1, Eqs. 6–9]
    "The key advantage here is that the 3D coordinates in X r ak come directly from the 3DGS model, ensuring strict scale consistency with the reconstructed scene. Consequently, the PnP solution inherently preserves the correct scale."

    Pose tracking uses the rendered pointmap X^r as the 3D reference; the pretrained pointmap contributes only 2D pixel correspondences (Eq. 7) and no 3D scale information. Thus 'correct scale' is defined as whatever scale the 3DGS map currently has. Any global scale error or slow scale drift in the Gaussian map is inherited by every PnP pose, and, because the mapping-side loss (Eq. 15) rescales the pretrained prior to match X^r, no component of the pipeline compares the map scale to an external metric after initialization. The tracking is scale-consistent with the map by construction; the paper's abstract-level claim of avoiding cumulative scale drift is not derivable from this anchoring alone and requires external scale validation that the paper does not report.

full rationale

The core reduction is in the mapping loop: Eq. 13 estimates σ′ from X^r, Eq. 14 uses the rescaled prior to replace points, and Eq. 15 supervises X^r with that same rescaled prior. Consequently the scale component of the geometric prior is conformed to the rendered map, making 'global scale consistency' a self-consistency of the map's own gauge rather than a prediction checked against an independent scale reference. The tracking module (Eqs. 6–9) similarly defines the 'correct scale' as the current 3DGS map scale, so it cannot by itself substantiate the headline claim that cumulative scale drift is avoided. This is partial circularity: the method is a coherent, empirically tested SLAM system, and its NVS and ATE numbers on Waymo/KITTI/DL3DV are external benchmarks, so the derivation is not wholly forced. The self-citations (notably [44]) are not load-bearing for the central construction. Still, the paper never reports a direct scale-drift metric or the trajectory-alignment protocol used for ATE, so the scale-consistency claim rests on the self-referential alignment loop. Score 6 reflects one or more central 'predictions' (scale consistency, preservation of correct scale) that reduce by construction to the map's own scale.

Assumptions & free parameters 9 free parameters · 4 assumptions · 0 invented entities

The scale of the map is inherited from a MASt3R pointmap initialization (Fig. 2 caption) and then maintained by scale factors computed from the map's own rendering. The method assumes static scenes, reliable MASt3R pointmaps, and accurate enough rendered pointmaps to serve as anchors. The hand-set hyperparameters in Algorithm 1 and Eq. 17 are free parameters that affect the central results.

free parameters (9)
  • Patch size P = 10
    Patch width for scale alignment (Algorithm 1); ablated on KITTI (Table 8), fixed at 10 for all datasets.
  • Patch selection thresholds delta_mu, delta_sigma = 0.3, 0.3
    Thresholds for selecting patches with similar distributions in Algorithm 1; chosen by hand.
  • Correct-point threshold epsilon_r = 0.1
    Threshold in Algorithm 1 for selecting points after normalization.
  • Scale alignment iteration limit max_iter = 3
    Iteration count for Algorithm 1.
  • Remedial fraction threshold tau = 0.01
    Minimum fraction of correct points before remedial scale factor is used.
  • Replacement threshold epsilon_m = 0.15
    Threshold in Eq. 14 for replacing incorrect points in rendered pointmap.
  • Loss weight alpha = 0.98
    Photometric-geometric loss weight in Eq. 17; text suggests 0.96 for confined scenes.
  • Isotropic regularization weight lambda_iso = 10
    Weight for Liso in Eq. 17.
  • Gaussian map initialization steps = 1000
    The 3D Gaussian map is initialized by optimizing MASt3R's pointmap for 1000 steps (Fig. 2 caption); this choice sets the initial map scale and density.
assumptions (4)
  • domain assumption MASt3R provides pointmaps with consistent cross-frame scale and reliable correspondences in outdoor scenes.
    Used in Eq. 7 to establish correspondences between keyframe and current frame; if the pretrained model's scale or correspondences are unreliable, the tracking chain breaks.
  • domain assumption The scene is static or nearly static; dynamic objects are ignored.
    Experiments select static scenes (Sec. 4.1), and Sec. 10 acknowledges the method cannot handle dynamic objects.
  • domain assumption The rendered 3DGS pointmap X^r is accurate enough to serve as a geometric anchor for pose estimation.
    Secs. 3.2.1 and 3.3.1 rely on X^r to provide 3D coordinates for PnP and scale fitting; if the Gaussian map has errors, the pose inherits them.
  • ad hoc to paper Patches with similar mean and standard deviation correspond to overlapping scene regions between X^r and X^p.
    Algorithm 1 selects patch candidates by |mu_r - mu_p| < delta_mu * mu_p and |sigma_r - sigma_p| < delta_sigma * sigma_p; this heuristic is not derived from any model.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Outdoor Monocular SLAM with Global Scale-Consistent 3D Gaussian Pointmaps." pith.science (2026). https://pith.science/paper/IA4BIUCD

@misc{pith2026250703737,
  author       = {Pith},
  title        = {Pith review of: Outdoor Monocular SLAM with Global Scale-Consistent 3D Gaussian Pointmaps},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/IA4BIUCD}},
  note         = {Machine review of arXiv:2507.03737}
}
read the original abstract

3D Gaussian Splatting (3DGS) has become a popular solution in SLAM due to its high-fidelity and real-time novel view synthesis performance. However, some previous 3DGS SLAM methods employ a differentiable rendering pipeline for tracking, lack geometric priors in outdoor scenes. Other approaches introduce separate tracking modules, but they accumulate errors with significant camera movement, leading to scale drift. To address these challenges, we propose a robust RGB-only outdoor 3DGS SLAM method: S3PO-GS. Technically, we establish a self-consistent tracking module anchored in the 3DGS pointmap, which avoids cumulative scale drift and achieves more precise and robust tracking with fewer iterations. Additionally, we design a patch-based pointmap dynamic mapping module, which introduces geometric priors while avoiding scale ambiguity. This significantly enhances tracking accuracy and the quality of scene reconstruction, making it particularly suitable for complex outdoor environments. Our experiments on the Waymo, KITTI, and DL3DV datasets demonstrate that S3PO-GS achieves state-of-the-art results in novel view synthesis and outperforms other 3DGS SLAM methods in tracking accuracy. Project page: https://3dagentworld.github.io/S3PO-GS/.

Figures

Figures reproduced from arXiv: 2507.03737 by the authors.

Figure 1
Figure 1. Localization and novel view synthesis results on KITTI. Our method S3PO-GS maintains robust tracking and high-quality novel view synthesis even in cases of large-angle turns. This is achieved through our self-consistent 3DGS pointmap tracking and the patch-based pointmap dynamic mapping module. Abstract 3D Gaussian Splatting (3DGS) has become a popular so￾lution in SLAM due to its high-fidelity and real-time novel v… view at source ↗
Figure 2
Figure 2. S3PO-GS pipeline for SLAM. The system begins by initializing a 3D Gaussian map (optimizing MASt3R’s pointmap for 1000 steps). For new input frame Tn, we rasterize the 3DGS pointmap of the adjacent keyframe Tak, match it with the input image, and establish 2D-3D correspondences to estimate scale self-consistent pose. The estimated pose is further refined using photometric loss. If Tn is selected as keyframe, we obtai… view at source ↗
Figure 3
Figure 3. Novel View Synthesis Results on Waymo (top three rows) and KITTI (bottom three rows) scenes, including Rendered RGB and Depth Maps. Our method produces high-fidelity images that capture intricate details of vehicles, streets, and buildings. The rendered depth maps are more accurate in regions with complex depth variations, such as tree branches and roadside vehicles. Method Waymo [36] KITTI [9] DL3DV [20] ATE↓ PSNR↑… view at source ↗
Figures from the paper (7 more)
Figure 4
Figure 4. Figure 4: Novel View Synthesis Results on DL3DV scenes, including Rendered RGB and Depth Maps. Our method renders clearer images with better reconstruction of details such as flowerbeds, eaves, and lanterns. The depth maps more accurately capture interwoven objects like columns …
Figure 5
Figure 5. Figure 5: Comparison of Tracking Trajectories with OpenGS￾SLAM and MonoGS. Under large viewpoint changes, MonoGS struggles to track, while OpenGS-SLAM exhibits instability. In contrast, our method achieves superior robustness. ods NeRF-SLAM [32] and NICER-SLAM [48], implicit enc…
Figure 6
Figure 6. Figure 6: Comparison with direct incorporation of pre-trained pointmap information in 3DGS-based SLAM (e.g., MonoGS). Scale drift leads to noticeable geometric blurring. further refines pose accuracy. Patch-based Pointmap Dynamic Mapping. As shown in Tab. 5, the absence of point…
Figure 7
Figure 7. Figure 7: Comparison of Tracking Trajectories with MonoGS and OpenGS-SLAM. 8.3. Ablation Study on Patch Size Tab. 8 shows that a patch size of 10–16 is optimal: larger patches introduce too many outliers, while smaller ones yield noisy statistics. 9. Additional Qualitative Resul…
Figure 8
Figure 8. Figure 8: Novel View Synthesis Results on Waymo, including Rendered RGB and Depth Maps. GlORIE-SLAM OpenGS-SLAM Ours Ground Truth PSNR:17.90 PSNR: 24.16 PSNR: 32.22 PSNR:16.93 PSNR: 27.51 PSNR: 33.27 MonoGS PSNR: 27.67 PSNR: 27.46 PSNR:18.90 PSNR: 24.69 PSNR: 30.34 PSNR:17.01 PS…
Figure 9
Figure 9. Figure 9: Novel View Synthesis Results on DL3DV, including rendered RGB and depth maps [PITH_FULL_IMAGE:figures/full_fig_p014_9.png]
Figure 10
Figure 10. Figure 10: Novel View Synthesis Results on KITTI, including Rendered RGB and Depth Maps [PITH_FULL_IMAGE:figures/full_fig_p015_10.png]

Discussion (0). Sign in to comment.

Forward citations

Cited by 1 Pith paper

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

  1. GeoGS-SLAM: Online Monocular Reconstruction Using Gaussian Splatting with Geometric Priors

    cs.RO 2026-07 conditional novelty 5.0 of 10

    An online monocular SLAM system that samples 3D Gaussians from RGB plus VGGT geometric priors and jointly optimizes poses and map with photometric and geometric losses plus loop closure, beating prior monocular 3DGS a...

Reference graph

Works this paper leans on

55 extracted references · 38 canonical work pages · cited by 1 Pith paper

  1. [1]

    Surf: Speeded up robust features

    Herbert Bay, Tinne Tuytelaars, and Luc Van Gool. Surf: Speeded up robust features. In Computer Vision–ECCV 2006: 9th European Conference on Computer Vision, Graz, Austria, May 7-13, 2006. Proceedings, Part I 9, pages 404–

  2. [2]

    Dsac - differentiable ransac for camera localization,

    Eric Brachmann, Alexander Krull, Sebastian Nowozin, Jamie Shotton, Frank Michel, Stefan Gumhold, and Carsten Rother. Dsac - differentiable ransac for camera localization,

  3. [3]

    Gomez Rodriguez, Jose M

    Carlos Campos, Richard Elvira, Juan J. Gomez Rodriguez, Jose M. M. Montiel, and Juan D. Tardos. Orb-slam3: An accurate open-source library for visual, visual–inertial, and multimap slam. IEEE Transactions on Robotics , 37(6): 1874–1890, 2021. 1, 2

  4. [4]

    A survey on 3d gaussian splatting

    Guikun Chen and Wenguan Wang. A survey on 3d gaussian splatting. arXiv preprint arXiv:2401.03890, 2024. 3

  5. [5]

    Direct sparse odometry

    Jakob Engel, Vladlen Koltun, and Daniel Cremers. Direct sparse odometry. IEEE transactions on pattern analysis and machine intelligence, 40(3):611–625, 2017. 5

  6. [6]

    Fischler and Robert C

    Martin A. Fischler and Robert C. Bolles. Random sample consensus: a paradigm for model fitting with applications to image analysis and automated cartography. Commun. ACM, 24(6):381–395, 1981. 4

  7. [7]

    Colmap-free 3d gaussian splat- ting

    Yang Fu, Sifei Liu, Amey Kulkarni, Jan Kautz, Alexei A Efros, and Xiaolong Wang. Colmap-free 3d gaussian splat- ting. In Proceedings of the IEEE/CVF Conference on Com- puter Vision and Pattern Recognition , pages 20796–20805,

  8. [8]

    Fastnerf: High-fidelity neu- ral rendering at 200fps

    Stephan J Garbin, Marek Kowalski, Matthew Johnson, Jamie Shotton, and Julien Valentin. Fastnerf: High-fidelity neu- ral rendering at 200fps. In Proceedings of the IEEE/CVF international conference on computer vision , pages 14346– 14355, 2021. 3

Show all 55 references
  1. [9]

    Vision meets robotics: The kitti dataset

    Andreas Geiger, Philip Lenz, Christoph Stiller, and Raquel Urtasun. Vision meets robotics: The kitti dataset. The in- ternational journal of robotics research, 32(11):1231–1237,

  2. [10]

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

    Jiarui Hu, Xianhao Chen, Boyin Feng, Guanglin Li, Liangjing Yang, Hujun Bao, Guofeng Zhang, and Zhaopeng Cui. Cg-slam: Efficient dense rgb-d slam in a consis- tent uncertainty-aware 3d gaussian field. arXiv preprint arXiv:2403.16095, 2024. 2, 3

  3. [11]

    Efficientnerf efficient neural radiance fields

    Tao Hu, Shu Liu, Yilun Chen, Tiancheng Shen, and Jiaya Jia. Efficientnerf efficient neural radiance fields. In Proceed- ings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12902–12911, 2022. 3

  4. [12]

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

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

  5. [13]

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

    Mohammad Mahdi Johari, Camilla Carta, and Franc ¸ois Fleuret. Eslam: Efficient dense slam system based on hybrid representation of signed distance fields. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 17408–17419, 2023. 3

  6. [14]

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

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

  7. [15]

    3d gaussian splatting for real-time radiance field rendering

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

  8. [16]

    Parallel tracking and map- ping for small ar workspaces

    Georg Klein and David Murray. Parallel tracking and map- ping for small ar workspaces. In 2007 6th IEEE and ACM international symposium on mixed and augmented reality , pages 225–234. IEEE, 2007. 2

  9. [17]

    Visual slam for autonomous ground vehicles

    Henning Lategahn, Andreas Geiger, and Bernd Kitt. Visual slam for autonomous ground vehicles. In2011 IEEE Interna- tional Conference on Robotics and Automation, pages 1732–

  10. [18]

    Ep n p: An accurate o (n) solution to the p n p problem

    Vincent Lepetit, Francesc Moreno-Noguer, and Pascal Fua. Ep n p: An accurate o (n) solution to the p n p problem. International journal of computer vision, 81:155–166, 2009. 4

  11. [19]

    Ground- ing image matching in 3d with mast3r

    Vincent Leroy, Yohann Cabon, and J´erˆome Revaud. Ground- ing image matching in 3d with mast3r. arXiv preprint arXiv:2406.09756, 2024. 3, 4, 5, 1, 2

  12. [20]

    Dl3dv-10k: A large-scale scene dataset for deep learning-based 3d vision

    Lu Ling, Yichen Sheng, Zhi Tu, Wentian Zhao, Cheng Xin, Kun Wan, Lantao Yu, Qianyu Guo, Zixun Yu, Yawen Lu, et al. Dl3dv-10k: A large-scale scene dataset for deep learning-based 3d vision. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , p...

  13. [21]

    Distinctive image features from scale- invariant keypoints

    David G Lowe. Distinctive image features from scale- invariant keypoints. International journal of computer vi- sion, 60:91–110, 2004. 2

  14. [22]

    Gaussian splatting slam

    Hidenobu Matsuki, Riku Murai, Paul HJ Kelly, and An- drew J Davison. Gaussian splatting slam. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pat- tern Recognition, pages 18039–18048, 2024. 2, 3, 5, 6, 7, 1

  15. [23]

    Srinivasan, Matthew Tancik, Jonathan T

    Ben Mildenhall, Pratul P. Srinivasan, Matthew Tancik, Jonathan T. Barron, Ravi Ramamoorthi, and Ren Ng. Nerf: Representing scenes as neural radiance fields for view syn- thesis, 2020. 2

  16. [24]

    Instant neural graphics primitives with a mul- tiresolution hash encoding

    Thomas M ¨uller, Alex Evans, Christoph Schied, and Alexan- der Keller. Instant neural graphics primitives with a mul- tiresolution hash encoding. ACM transactions on graphics (TOG), 41(4):1–15, 2022. 3

  17. [25]

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

    Raul Mur-Artal and Juan D Tard ´os. Orb-slam2: An open- source slam system for monocular, stereo, and rgb-d cam- eras. IEEE transactions on robotics, 33(5):1255–1262, 2017. 1, 2

  18. [26]

    Orb-slam: a versatile and accurate monocular slam system

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

  19. [27]

    Mast3r-slam: Real-time dense slam with 3d reconstruction priors

    Riku Murai, Eric Dexheimer, and Andrew J Davison. Mast3r-slam: Real-time dense slam with 3d reconstruction priors. In Proceedings of the Computer Vision and Pattern Recognition Conference, pages 16695–16705, 2025. 2

  20. [28]

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

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

  21. [29]

    Dtam: Dense tracking and mapping in real-time

    Richard A Newcombe, Steven J Lovegrove, and Andrew J Davison. Dtam: Dense tracking and mapping in real-time. In 2011 international conference on computer vision , pages 2320–2327. IEEE, 2011. 1, 2

  22. [30]

    A framework for the volumetric integration of depth images

    Victor Adrian Prisacariu, Olaf K ¨ahler, Ming Ming Cheng, Carl Yuheng Ren, Julien Valentin, Philip HS Torr, Ian D Reid, and David W Murray. A framework for the volumetric integration of depth images. arXiv preprint arXiv:1410.0925, 2014. 2

  23. [31]

    Sacreg: Scene-agnostic co- ordinate regression for visual localization, 2023

    Jerome Revaud, Yohann Cabon, Romain Br ´egier, JongMin Lee, and Philippe Weinzaepfel. Sacreg: Scene-agnostic co- ordinate regression for visual localization, 2023. 3

  24. [32]

    Leonard, and Luca Carlone

    Antoni Rosinol, John J. Leonard, and Luca Carlone. Nerf- slam: Real-time dense monocular slam with neural radiance fields, 2022. 3, 6, 7

  25. [33]

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

    Erik Sandstr ¨om, Keisuke Tateno, Michael Oechsle, Michael Niemeyer, Luc Van Gool, Martin R Oswald, and Federico Tombari. Splat-slam: Globally optimized rgb-only slam with 3d gaussians. arXiv preprint arXiv:2405.16544, 2024. 2

  26. [34]

    Os- wald

    Erik Sandstr ¨om, Yue Li, Luc Van Gool, and Martin R. Os- wald. Point-slam: Dense neural point cloud-based slam,

  27. [35]

    imap: Implicit mapping and positioning in real-time

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

  28. [36]

    Scalability in perception for autonomous driving: Waymo open dataset

    Pei Sun, Henrik Kretzschmar, Xerxes Dotiwalla, Aurelien Chouard, Vijaysai Patnaik, Paul Tsui, James Guo, Yin Zhou, Yuning Chai, Benjamin Caine, Vijay Vasudevan, Wei Han, Jiquan Ngiam, Hang Zhao, Aleksei Timofeev, Scott Et- tinger, Maxim Krivokon, Amy Gao, Aditya Joshi, Yu Zhan...

  29. [37]

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

    Hengyi Wang, Jingwen Wang, and Lourdes Agapito. Co- slam: Joint coordinate and sparse parametric encodings for neural real-time slam. In Proceedings of the IEEE/CVF Con- ference on Computer Vision and Pattern Recognition, pages 13293–13302, 2023. 3

  30. [38]

    Dust3r: Geometric 3d vi- sion made easy

    Shuzhe Wang, Vincent Leroy, Yohann Cabon, Boris Chidlovskii, and Jerome Revaud. Dust3r: Geometric 3d vi- sion made easy. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 20697– 20709, 2024. 3, 4

  31. [39]

    Bovik, H.R

    Zhou Wang, A.C. Bovik, H.R. Sheikh, and E.P. Simoncelli. Image quality assessment: from error visibility to structural similarity. IEEE Transactions on Image Processing , 13(4): 600–612, 2004. 6

  32. [40]

    Elasticfusion: Dense slam without a pose graph

    Thomas Whelan, Stefan Leutenegger, Renato Moreno, Ben Glocker, and Andrew Davison. Elasticfusion: Dense slam without a pose graph. 2015. 2

  33. [41]

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

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

  34. [42]

    Sanet: Scene agnostic net- work for camera localization

    Luwei Yang, Ziqian Bai, Chengzhou Tang, Honghua Li, Ya- sutaka Furukawa, and Ping Tan. Sanet: Scene agnostic net- work for camera localization. 2019 IEEE/CVF International Conference on Computer Vision (ICCV), pages 42–51, 2019. 3

  35. [43]

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

    Xingrui Yang, Hai Li, Hongjia Zhai, Yuhang Ming, Yuqian Liu, and Guofeng Zhang. V ox-fusion: Dense tracking and mapping with voxel-based neural implicit representation. In 2022 IEEE International Symposium on Mixed and Aug- mented Reality (ISMAR), pages 499–507. IEEE, 2022. 3

  36. [44]

    Rgb-only gaussian splatting slam for unbounded outdoor scenes

    Sicheng Yu, Chong Cheng, Yifan Zhou, Xiaojun Yang, and Hao Wang. Rgb-only gaussian splatting slam for unbounded outdoor scenes. arXiv preprint arXiv:2502.15633, 2025. 2, 3, 4, 5, 6, 7

  37. [45]

    Glorie-slam: Glob- ally optimized rgb-only implicit encoding point cloud slam

    Ganlin Zhang, Erik Sandstr ¨om, Youmin Zhang, Manthan Pa- tel, Luc Van Gool, and Martin R Oswald. Glorie-slam: Glob- ally optimized rgb-only implicit encoding point cloud slam. arXiv preprint arXiv:2403.19549, 2024. 1, 2, 5, 6, 7

  38. [46]

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

    Pengcheng Zhu, Yaoming Zhuang, Baoquan Chen, Li Li, Chengdong Wu, and Zhanlin Liu. Mgs-slam: Monocular sparse tracking and gaussian mapping with depth smooth regularization. arXiv preprint arXiv:2405.06241 , 2024. 2, 3, 5

  39. [47]

    Nice-slam: Neural implicit scalable encoding for slam

    Zihan Zhu, Songyou Peng, Viktor Larsson, Weiwei Xu, Hu- jun Bao, Zhaopeng Cui, Martin R Oswald, and Marc Polle- feys. Nice-slam: Neural implicit scalable encoding for slam. In Proceedings of the IEEE/CVF conference on computer vi- sion and pattern recognition, pages 12786–1279...

  40. [48]

    Nicer-slam: Neural implicit scene encoding for rgb slam

    Zihan Zhu, Songyou Peng, Viktor Larsson, Zhaopeng Cui, Martin R Oswald, Andreas Geiger, and Marc Pollefeys. Nicer-slam: Neural implicit scene encoding for rgb slam. In 2024 International Conference on 3D Vision (3DV) , pages 42–52. IEEE, 2024. 3, 6, 7 Outdoor Monocular SLAM wi...

  41. [49]

    We also include additional experiments on KITTI, with runtime, memory, and patch size analysis

    Overview This supplementary material provides implementation de- tails for keyframe management, patch-based scale align- ment, pointmap replacement, and Gaussian map optimiza- tion modules. We also include additional experiments on KITTI, with runtime, memory, and patch size a...

  42. [50]

    correct points

    Implementation Details 7.1. Keyframe Management As described in Section 3.3.3, we joinly refine camera poses and the Gaussian map within a local keyframe window W. A well-designed keyframe selection strategy must en- sure sufficient viewpoint coverage while avoiding redun- dan...

  43. [51]

    We also analyzed runtime and memory con- sumption, and performed an ablation study on the patch size used in Algorithm 1

    Additional Experiments We conducted additional experiments on the KITTI-07 se- quence, including further comparisons with CF-3DGS [7] , MASt3R-SLAM [19], DROID-SLAM [27], and Splat- SLAM [33]. We also analyzed runtime and memory con- sumption, and performed an ablation study o...

  44. [52]

    Figures 8 to 10 shows additional novel view synthesis re- sults in the Waymo, DL3DV , and KITTI datasets

    Additional Qualitative Results Figure 7 presents additional trajectory comparisons, further highlighting the robustness of our method in location under challenging outdoor environments. Figures 8 to 10 shows additional novel view synthesis re- sults in the Waymo, DL3DV , and K...

  45. [53]

    Limitations and Future Works

  46. [54]

    Monocular RGB-only SLAM for outdoor envi- ronments with dynamic objects remains a highly inter- esting and challenging problem

    Our method cannot handle dynamic objects in outdoor scenes. Monocular RGB-only SLAM for outdoor envi- ronments with dynamic objects remains a highly inter- esting and challenging problem

  47. [55]

    While their inclusion would benefit long-sequence SLAM, it also introduces challenges related to training time and memory consumption

    Our method does not incorporate loop closure or global BA. While their inclusion would benefit long-sequence SLAM, it also introduces challenges related to training time and memory consumption. GlORIE-SLAM OpenGS-SLAM Ours Ground Truth PSNR: 21.40 PSNR: 24.48 PSNR: 22.38 PSNR:...

Pith tools

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