Pith. sign in

REVIEW 3 major objections 4 minor 34 references

GLAM-SLAM claims that a monocular Gaussian-splatting SLAM system can sustain photorealistic, long-sequence outdoor mapping at 10–20 FPS by seeding Gaussian anchors with epipolar-filtered optical flow and using localized MLPs per scene regio

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

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

A real-time Gaussian-splatting SLAM system densifies sparse ORB-SLAM2 maps with epipolar-filtered optical flow and splits long routes into localized MLP regions, completing 4500-frame outdoor sequences the leading systems cannot.

T0 review reviewed 2026-08-01 challenge →

load-bearing objection A genuinely useful, honestly-reported 3DGS-SLAM system that scales to long outdoor sequences; the headline '15%' overstates and the flow-seeding hyperparameters need reporting, but the core contribution is real. the 3 major comments →

arxiv 2607.21416 v1 pith:ESIZMNR5 submitted 2026-07-23 cs.RO cs.CV

GLAM-SLAM: Real-time Gaussian Large-scale Mapping via Flow Densification and Spatial Decomposition

Panagiotis Mermigkas , Argyris Manetas , Petros Maragos This is my paper
classification cs.RO cs.CV
keywords Gaussian Splatting SLAMreal-time dense mappinglarge-scale outdoor reconstructionoptical flow densificationepipolar geometrylocalized MLPsanchor gridmonocular SLAM
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

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

The reading

The paper tries to establish that photorealistic dense mapping of long outdoor monocular drives can be done in real time, on a single 32 GB GPU, by decoupling tracking from Gaussian-splatting mapping. It argues that a sparse feature-based frontend can supply enough structure if the Gaussian map is seeded with extra geometric anchors obtained from dense optical flow filtered by epipolar constraints. It further argues that splitting the scene into regions, each governed by its own small MLP, prevents the appearance and illumination drift that a single global network suffers on long sequences. The reported evidence: across KITTI, Oxford RobotCar and Málaga, the system achieves 10–20 FPS with the lowest peak GPU memory among compared real-time systems, completes sequences that competing Gaussian-SLAM systems cannot finish due to memory exhaustion, and averages 17.3 dB PSNR on KITTI, about 15% above the next-best performer before offline refinement. A sympathetic reading accepts the central trade: slightly noisy flow correspondences are safe for mapping even though they would be unsuitable for tracking.

Core claim

On its own terms, the paper's central claim is that two mechanisms—flow-based densification and localized MLP initialization—turn a sparse, feature-based SLAM frontend into a dense photorealistic mapper that scales to unbounded outdoor sequences. The flow module computes dense correspondences between keyframes with a lightweight optical-flow network, keeps only matches whose epipolar residual is below a threshold, triangulates them with the tracked poses, and voxelizes the union of tracked landmarks and triangulated flow points into Gaussian anchors. The localization module partitions the trajectory into regions triggered by large rotation changes and assigns each region an independent MLP t

What carries the argument

Two coupled mechanisms carry the argument. (1) Flow-guided densification: a dense optical-flow model supplies pixel correspondences between keyframes once every seven keyframes; correspondences are kept only if they satisfy the epipolar residual condition |x'^T F x| < τ, then triangulated with the tracked poses and merged with tracked landmarks into a voxelized anchor grid. This supplies the dense geometric prior that Gaussian optimization needs, at a fixed GPU cost for the flow model. (2) Localized MLP initialization: the mapped environment is partitioned into regions, and each region has its own parameter set; an indicator function activates only the MLP of the region containing the camera

Load-bearing premise

The load-bearing premise is that flow matches passing a single per-image epipolar residual test are geometrically trustworthy enough to seed Gaussian anchors; the threshold value and sample count are not reported, and the frontend's poses can drift badly on some sequences, so a wrong seed that passes the filter can poison the map before the photometric loss can repair it.

What would settle it

Triangulate the flow correspondences on KITTI Seq.08 and compare their depths to LiDAR ground truth at the same keyframes. If a substantial fraction, say more than 20%, of correspondences that pass the epipolar residual condition have depth error above 50%, the epipolar filter is not sufficient and the reported quality gain should not be attributed to geometrically correct densification; equivalently, disabling flow densification on the high-ATE sequences should leave PSNR nearly unchanged if the filter had already lost all informative seeds.

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

If this is right

  • Long-horizon mapping becomes practical: whole KITTI sequences of 4,000+ frames complete without out-of-memory failure, where the compared real-time systems stop at 2,000–3,200 frames.
  • Real-time quality no longer requires an offline color-refinement pass; the pre-refinement map already beats the next-best system on photometric metrics.
  • Decoupling means a mapping-only flow network can improve density without endangering tracking accuracy, and tracking stays CPU-only at 10 FPS.
  • The two contributions are complementary: the 1,000-frame KITTI ablation shows PSNR rising from 17.770 at baseline to 18.320 with flow alone, 18.388 with MLPs alone, and 18.800 with both.
  • If the claims hold, a single 32 GB GPU is sufficient for photorealistic online maps of kilometer-scale outdoor trajectories, removing a barrier to deployment.

Where Pith is reading between the lines

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

  • Inference: because the epipolar filter is the only geometric gate on flow seeds, the system's worst-case robustness is set by frontend pose drift and dynamic scenes; high-ATE sequences are exactly where the densification could seed false geometry, a stress test the paper's averaged metrics do not isolate.
  • Inference: the turn-detection partitioning rule is a heuristic; a principled online partition criterion based on anchor coverage or appearance change could tighten region boundaries and reduce cross-boundary photometric fluctuation.
  • Inference: the flow model's fixed GPU footprint and fixed 7-keyframe step suggest an adaptive schedule—run flow only in sparse anchor regions—could cut memory further and free iterations for Gaussian optimization.
  • Inference: the same seeding recipe, flow plus epipolar filter plus triangulation, is a drop-in densifier for other sparse SLAM frontends, so the contribution may be reusable beyond the presented system.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

3 major / 4 minor

Summary. GLAM-SLAM is a monocular Gaussian-splatting SLAM system with a decoupled architecture: ORB-SLAM2 provides camera tracking on the CPU, while a GPU-based Gaussian mapper builds a Scaffold-GS-style anchor grid. The paper makes two main technical contributions: (i) a flow-densification module that uses optical flow and an epipolar residual test to add geometric anchors in regions where ORB-SLAM2 points are sparse, and (ii) a spatial-decomposition strategy that partitions the scene and assigns localized MLP parameter sets to different regions. The system is evaluated on KITTI Odometry, Oxford RobotCar, Málaga, and a self-captured parking sequence, reporting real-time frame rates, lower GPU memory than comparison methods, and improved PSNR/SSIM/LPIPS relative to PhotoSLAM and to GigaSLAM before its offline post-optimization. The paper also includes ablations on the two contributions and an ATE evaluation showing that ORB-SLAM2 tracking can drift substantially on some sequences.

Significance. If the results hold, GLAM-SLAM is a practically valuable system: it is one of the few Gaussian-splatting SLAM systems that operates in real time on long outdoor sequences with bounded GPU memory, and the decoupling of tracking from mapping is a sensible design choice. The paper's strengths include a public code release, clean ablations isolating each contribution (Table V shows independent gains for flow and MLP, and a further gain when combined), and evaluation on multiple outdoor benchmarks including held-out Oxford and Málaga sequences. The reported real-time FPS and peak-memory numbers, including completion of sequences where PhotoSLAM and GigaSLAM exhaust memory, are useful engineering evidence. The main caveat is that the flow-densification step relies on an epipolar-only correspondence filter whose parameters are not reported and whose robustness on long, drift-prone sequences is not directly validated; this is the key correctness risk for the scaling and quality claims.

major comments (3)
  1. [§III-C, Eq. (5)] The flow-densification filter retains correspondences only when |x'^T F x| < τ, with N random samples; τ and N are never reported. An epipolar residual is a necessary condition, not a sufficient one: any point on the epipolar line can satisfy it, so along-epipolar flow errors, moving vehicles, and motion blur can produce accepted wrong matches. Because anchors are fixed after initialization and the photometric loss only optimizes Gaussians spawned from them, a poisoned anchor cannot be repaired. This is especially concerning on long sequences where Table IV shows ORB-SLAM2 poses can drift badly (Seq.08 ATE 46.07 m, Seq.02 17.61 m); under such drift the fundamental matrix in Eq. (5) is itself unreliable. Please report τ and N, add a multi-view or depth-consistency check, and provide a sensitivity analysis or an ablation that varies these values.
  2. [§IV-F / Table V vs. §IV-C / Table I] The ablations in Table V are limited to 1000-frame prefixes, whereas the full-sequence results in Tables I and III always have flow densification and localized MLPs enabled. There is no full-length flow-off control, so the long-sequence behavior of the proposed densification is not tested against the baseline on the exact regimes where the scalability claim is made. The failure mode identified above—wrong flow correspondences passing the epipolar test—is likely to accumulate over thousands of frames, especially on sequences with high ATE. Please add a full-length flow-off baseline (or at least report per-sequence comparisons with and without flow on the long KITTI sequences Seq.00, Seq.02, and Seq.08).
  3. [§IV-C, Tables I and III] The comparison with PhotoSLAM and GigaSLAM is partially unfair because those baselines are truncated by out-of-memory failures (2000 and 3200 frames, respectively) while GLAM-SLAM runs the full sequence. Reporting per-sequence averages over different numbers of frames can bias the comparison if later frames are systematically harder or easier. The headline '15% improvement over the second-best performer' is therefore not fully supported. Please include length-matched comparisons on the common prefixes, or show per-frame quality curves over the full trajectory, so that the advantage is not confounded by sequence length.
minor comments (4)
  1. [Table I] The formatting of Table I is broken in places: values and '×' markers run together (e.g., '11.161× 18.33613.613 ×'), making it hard to determine which entry belongs to which sequence. Please reformat with separated columns.
  2. [§IV-F] The text says the ablation excludes Seq.00 because ORB-SLAM2 loses tracking, but Table V lists 1000 frames for Seq.00 and marks Seq.01 as failed. This contradiction should be corrected.
  3. [§III-C / §III-D] Several free parameters are unreported: the sampled correspondence count N, the epipolar threshold τ, the turn-detection threshold, and the number of regions M. Even if defaults come from Scaffold-GS or ORB-SLAM2, the effective values should be stated.
  4. [Abstract / §IV-C] The abstract's '15% improvement' is not directly traceable to one table; the average PSNR gains are 11.6% on KITTI, 26.6% on Oxford, and 15.7% on Málaga. Please qualify the claim or present an aggregate that is explicitly defined.

Circularity Check

0 steps flagged

No load-bearing derivation reduces to its inputs; Eq. 4 and Eq. 6 are constructive definitions validated by ablations, and no self-citation chain is used.

full rationale

The paper's central claims are empirically evaluated against external baselines (KITTI, Oxford, Málaga) and its two main contributions are ablated independently (Table V). Eq. 4 is a voxelization of fused ORB-SLAM2 and flow points, which is a representation construction, not a prediction derived from itself. Eq. 6 is a piecewise definition of localized MLP outputs; calling it a 'spatial inductive bias' is a description of the architecture, not a circular derivation. The flow-densification module filters correspondences using the epipolar residual of Eq. 5; this is a correctness/robustness concern (unreported τ and N, potential pose drift, e.g., ATE 46.07m on Seq.08), but the filter output feeds the map rather than being claimed as a prediction of the metrics. There are no self-citations in the reference list and no uniqueness theorem imported from the authors' prior work. The '15% improvement' claim is a summary of the reported table numbers, not a fitted parameter renamed as a prediction. Therefore no circular step can be identified under the specified criteria; score 0.

Axiom & Free-Parameter Ledger

6 free parameters · 5 axioms · 1 invented entities

The central claim rests on roughly six tuned-but-mostly-unreported hyperparameters (ε, τ, N, flow step, sampling weights, M/turn threshold) and five assumptions, the most fragile being that epipolar-filtered flow points are valid 3D seeds and that ORB-SLAM2 poses are good enough to serve as triangulation ground truth. The system introduces no physical entities; the localized MLP sets are architectural. The representation (anchor grid) and tracking (ORB-SLAM2) are inherited from prior work, which keeps the added-axiom burden moderate for a systems paper.

free parameters (6)
  • Anchor voxel size ε (Eq. 4) = 0.001 (Scaffold-GS default, §IV-B)
    Sets anchor grid resolution. The paper states larger values collapse initialization points and diminish the densification benefit, so results are sensitive to this choice.
  • Epipolar residual threshold τ (Eq. 5) = not reported
    Determines which flow matches become anchors; directly controls seeding density and accuracy. No value given anywhere in the paper.
  • Sampled correspondence count N (Eq. 5) = not reported
    Random subsample per keyframe for epipolar evaluation; affects computation, seed coverage, and run-to-run variance since sampling is stochastic.
  • Optical flow schedule step = 7 keyframes (§IV-B)
    Sparsity schedule for flow densification; chosen 'to prevent diminishing returns' without ablation.
  • Training keyframe sampling (k, p) = k=25, p(most recent)=0.7 (§IV-B)
    Sampling distribution used in every Gaussian optimization iteration to balance forgetting and recency; not ablated.
  • Region count M / turn-detection threshold = not reported
    Number of localized MLPs and the trigger sensitivity for partitioning (Eq. 6); the entire spatial-decomposition contribution hinges on this unstated heuristic.
axioms (5)
  • standard math Calibrated camera geometry: F = [e2]× P2 P1+ fully captures the epipolar constraint between consecutive keyframes (Eq. 5).
    Standard epipolar geometry; depends on known intrinsics and accurate ORB-SLAM2 poses.
  • domain assumption ORB-SLAM2 poses are accurate enough to serve as ground truth for triangulating flow points and placing anchors.
    Used throughout §III-C and §III-E. Violated on Seq.01 (tracking failed) and weakly on Seq.02/Seq.08 (ATE 17.6 m/46.1 m in Table IV), where the map quality must rely on local consistency rather than global pose accuracy.
  • domain assumption Flow matches passing the residual x'^T F x < τ are correct in 3D; no multi-view or depth verification and no dynamic-object masking.
    §III-C defines the only filter. Urban traffic contains moving objects for which the epipolar residual is not a validity test; dynamic content can seed false anchors.
  • domain assumption Scaffold-GS anchor-grid plus MLP decoding is an adequate base representation for long outdoor sequences.
    Adopted from [11] (§III-A) and used as the common baseline in ablations; the MLP ablation exists precisely because the single-MLP variant degrades in outdoor conditions.
  • domain assumption Absence of cross-region consistency in localized MLPs does not perceptually degrade rendering.
    Stated in §IV-F ('explicit cross-partition consistency is not enforced... do not perceptually degrade the rendering quality') with no quantitative boundary evaluation or metric.
invented entities (1)
  • Region-specific local MLP parameter sets Θ = {θ_1, ..., θ_M} no independent evidence
    purpose: Decode Gaussian attributes (position, opacity, SH color) locally per spatial region to handle illumination, scale, and surface variability across long routes (Eq. 6).
    Standard neural-network components, not a conjectured physical entity. Their contribution is shown only by the in-domain ablation (Table V); there is no formal guarantee of boundary consistency and the paper admits boundaries are unconstrained.

reviewed 2026-08-01 · how reviews work

0 comments
Cite this review

Pith. "Pith review of GLAM-SLAM: Real-time Gaussian Large-scale Mapping via Flow Densification and Spatial Decomposition." pith.science (2026). https://pith.science/paper/ESIZMNR5

@misc{pith2026260721416,
  author       = {Pith},
  title        = {Pith review of: GLAM-SLAM: Real-time Gaussian Large-scale Mapping via Flow Densification and Spatial Decomposition},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/ESIZMNR5}},
  note         = {Machine review of arXiv:2607.21416}
}
Share X Bluesky LinkedIn Reddit HN
read the original abstract

Existing Gaussian-splatting-based monocular Simultaneous Localization and Mapping (SLAM) systems are either tailored to short sequences, are not real-time, or suffer from prohibitive GPU memory requirements, limiting their applicability in realistic, long-horizon scenarios. To address this, we present GLAM-SLAM, a real-time, decoupled Gaussian-splatting SLAM system designed for large-scale outdoor scenes. We ensure lightweight tracking using a robust, feature-based SLAM frontend, while for mapping, we adopt a structured, sparse anchor grid representation that ensures scalable operation and maintains scene coherence across long-term sequences. To satisfy the dense initialization requirements of 3D Gaussian Splatting (3DGS), we introduce a geometry-based flow-densification anchoring strategy using epipolar constraints. Furthermore, by treating mapping as a multi-scene problem, we propose a scene-partitioning strategy that introduces a strong spatial inductive bias via MLP initializations to generate localized Gaussians. We evaluate our system on the challenging, long-sequence KITTI Odometry, Oxford RobotCar, and M'alaga datasets. Extensive ablations and comparisons demonstrate a 15% improvement in reconstruction quality over the second-best performer, while maintaining real-time performance and the ability to scale to longer sequences. Code is publicly available for the benefit of the community.

Figures

Figures reproduced from arXiv: 2607.21416 by Argyris Manetas, Panagiotis Mermigkas, Petros Maragos.

Figure 1
Figure 1. Figure 1: Novel view synthesis from unseen trajectories: These renderings, generated from camera poses not present in the KITTI (top) and Oxford RobotCar (bottom) datasets, demonstrate the strong generalization and geo￾metric consistency of our 3DGS-based map without any post-optimization. and dense geometric alignment, respectively. These methods offered a paradigm shift from representing the world as a set of dist… view at source ↗
Figure 2
Figure 2. Figure 2: GLAM (Gaussian LArge-scale Mapping)-SLAM system overview: Our system follows a decoupled design with an ORB-SLAM2 thread serving as the real-time front-end and a Gaussian Mapper thread that constructs its anchor grid by integrating sparse ORB-SLAM2 points with flow-densified points. Training is guided by a combined photometric and structural similarity objective as defined in [17]: L = λL1 + (1 − λ)LSSIM (… view at source ↗
Figure 3
Figure 3. Figure 3: Qualitative rendering comparison of our contributions on KITTI Odometry Seq. 00: The lack of structured anchors derived from ORB￾SLAM2 features in the left side of the scene leads to blurry localized renderings in our baseline implementation (c). Visual fidelity in this area is greatly improved when employing our individual contributions (d, e) and their combination (f). C. Flow Densification Sparse initia… view at source ↗
Figure 4
Figure 4. Figure 4: Qualitative rendering comparison on KITTI Odometry Seq. 08: Rows correspond to frames 500, 2000, and 4000. PhotoSLAM (a) exhibits significant rendering degradation over time. PhotoSLAM (a) and GigaSLAM (b) terminate due to memory exhaustion after 2000 and 3200 frames, respectively. MonoGS fails on this sequence and is not included. In contrast, our method maintains consistent rendering quality throughout a… view at source ↗
Figure 5
Figure 5. Figure 5: Trajectory comparison on KITTI Seqs. 00, 05: GLAM-SLAM is the only method that completes the entire 4541 and 2761-frame sequences. TABLE IV Absolute Trajectory Error (ATE RMSE [m]) on the KITTI Odometry Dataset. ×: Due to Out-of-Memory limitations, PhotoSLAM reports results for 2000 frames and GigaSLAM for 3200 frames. Method 00 01 02 03 04 05 06 07 08 09 10 Total frames 4541 1101 4661 801 271 2761 1101 11… view at source ↗
Figure 6
Figure 6. Figure 6: Per-keyframe quality analysis along the X–Z trajectory: We visualize absolute PSNR (first plot) and relative improvement ∆PSNR (remaining plots) for the ablation on KITTI Seq. 00 (first 1000 frames). The average PSNR and ∆PSNR values reported match those presented in Table V. F. Ablation Studies We evaluate our main contributions through a detailed ablation study of up to 1000 frames of each KITTI sequence… view at source ↗
Figure 7
Figure 7. Figure 7: Qualitative evaluation on the Parking sequence: (a) The data acquisition UGV. (b)–(f) Comparative rendering results. TABLE VI Photometric reconstruction quality results on the Parking sequence. Ablations → Baseline OF MLP OF+MLP GigaSLAM PhotoSLAM MonoGS Total frames 2200 2200 2200 2200 2200 2200 2200 PSNR ↑ 23.428 23.648 23.459 23.719 19.149 21.066 13.351 SSIM ↑ 0.963 0.964 0.963 0.964 0.911 0.612 0.492 L… view at source ↗

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

34 extracted references · 3 linked inside Pith

  1. [1]

    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 Transactions on Robotics, vol. 33, no. 5, pp. 1255–1262, Oct. 2017

    2017

  2. [2]

    KinectFusion: 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, “KinectFusion: Real-Time Dense Surface Mapping and Tracking,” in Proc. IEEE Int. Symp. on Mixed and Augmented Reality, 2011

    2011

  3. [3]

    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,” inProc. IEEE/CVF Conf. on Computer Vision and Pattern Recognition, 2022

    2022

  4. [4]

    Point-SLAM: Dense Neural Point Cloud-based SLAM,

    E. Sandstr ¨om, Y . Li, L. Van Gool, and M. R. Oswald, “Point-SLAM: Dense Neural Point Cloud-based SLAM,” inProc. IEEE/CVF Int. Conf. on Computer Vision, 2023

    2023

  5. [5]

    Gaussian Splatting SLAM,

    H. Matsuki, R. Murai, P. H. Kelly, and A. J. Davison, “Gaussian Splatting SLAM,” inProc. IEEE/CVF Conf. on Computer Vision and Pattern Recognition, 2024

    2024

  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,” inProc. IEEE/CVF Conf. on Computer Vision and Pattern Recognition, 2024

    2024

  7. [7]

    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,” inProc. IEEE/CVF Conf. on Computer Vision and Pattern Recognition, 2024

    2024

  8. [8]

    LSD-SLAM: Large-Scale Direct Monocular SLAM,

    J. Engel, T. Sch ¨ops, and D. Cremers, “LSD-SLAM: Large-Scale Direct Monocular SLAM,” inProc. European Conf. on Computer Vision, 2014

    2014

  9. [9]

    Relax- ing Accurate Initialization Constraint for 3D Gaussian Splatting,

    J. Jung, J. Han, H. An, J. Kang, S. Park, and S. Kim, “Relax- ing Accurate Initialization Constraint for 3D Gaussian Splatting,” arXiv:2403.09413, 2024

    Pith/arXiv arXiv 2024

  10. [10]

    Liberated-GS: 3D Gaussian Splatting Independent from SfM Point Clouds,

    W. Pan, X. Zhang, H. Zhai, X. Xiang, H. Jiang, and G. Zhang, “Liberated-GS: 3D Gaussian Splatting Independent from SfM Point Clouds,” inProc. IEEE/CVF Int. Conf. on Computer Vision, 2025

    2025

  11. [11]

    Scaffold-GS: Structured 3D Gaussians for View-Adaptive Render- ing,

    T. Lu, M. Yu, L. Xu, Y . Xiangli, L. Wang, D. Lin, and B. Dai, “Scaffold-GS: Structured 3D Gaussians for View-Adaptive Render- ing,” inProc. IEEE/CVF Conf. on Computer Vision and Pattern Recognition, 2024

    2024

  12. [12]

    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 Proc. Robotics: Science and Systems, 2015

    2015

  13. [13]

    Kintinuous: Spatially Extended KinectFusion,

    T. Whelan, J. McDonald, M. Kaess, M. Fallon, H. Johannsson, and J. J. Leonard, “Kintinuous: Spatially Extended KinectFusion,” inProc. RSS Workshop on RGB-D: Advanced Reasoning with Depth Cameras, 2012

    2012

  14. [14]

    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, Dec. 2021

    2021

  15. [15]

    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,” inProc. IEEE/CVF Int. Conf. on Computer Vision, 2021

    2021

  16. [16]

    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,” inProc. IEEE/CVF Conf. on Computer Vision and Pattern Recognition, 2024

    2024

  17. [17]

    3D Gaussian Splatting for Real-Time Radiance Field Rendering,

    B. Kerbl, G. Kopanas, T. Leimkuehler, and G. Drettakis, “3D Gaussian Splatting for Real-Time Radiance Field Rendering,”ACM Transac- tions on Graphics, vol. 42, no. 4, Jul. 2023

    2023

  18. [18]

    A Survey on 3D Gaussian Splatting,

    G. Chen and W. Wang, “A Survey on 3D Gaussian Splatting,” arXiv:2401.03890, 2024

    Pith/arXiv arXiv 2024

  19. [19]

    2D Gaussian Splatting for Geometrically Accurate Radiance Fields,

    B. Huang, Z. Yu, A. Chen, A. Geiger, and S. Gao, “2D Gaussian Splatting for Geometrically Accurate Radiance Fields,” inProc. ACM SIGGRAPH Conf., 2024

    2024

  20. [20]

    Globally Consistent RGB-D SLAM with 2D Gaussian Splatting,

    X. Zhong, Y . Pan, L. Jin, M. Popovi ´c, J. Behley, and C. Stachniss, “Globally Consistent RGB-D SLAM with 2D Gaussian Splatting,” IEEE Transactions on Robotics, vol. 42, pp. 2360–2380, May 2026

    2026

  21. [21]

    VINGS- Mono: Visual-Inertial Gaussian Splatting Monocular SLAM in Large Scenes,

    K. Wu, Z. Zhang, M. Tie, Z. Ai, Z. Gan, and W. Ding, “VINGS- Mono: Visual-Inertial Gaussian Splatting Monocular SLAM in Large Scenes,”IEEE Transactions on Robotics, vol. 41, pp. 5912–5931, Sep. 2025

    2025

  22. [22]

    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 Proc. IEEE/CVF Conf. on Computer Vision and Pattern Recognition, 2025

    2025

  23. [23]

    RTG-SLAM: Real-time 3D Reconstruction at Scale using Gaussian Splatting,

    Z. Peng, T. Shao, Y . Liu, J. Zhou, Y . Yang, J. Wang, and K. Zhou, “RTG-SLAM: Real-time 3D Reconstruction at Scale using Gaussian Splatting,” inProc. ACM SIGGRAPH Conf., 2024

    2024

  24. [24]

    GSFusion: Online RGB-D Mapping Where Gaussian Splatting Meets TSDF Fusion,

    J. Wei and S. Leutenegger, “GSFusion: Online RGB-D Mapping Where Gaussian Splatting Meets TSDF Fusion,”IEEE Robotics and Automation Letters, vol. 9, no. 12, pp. 11 865–11 872, Dec. 2024

    2024

  25. [25]

    VGGT-Long: Chunk it, Loop it, Align it–Pushing VGGT’s Limits on Kilometer-scale Long RGB Sequences,

    K. Deng, Z. Ti, J. Xu, J. Yang, and J. Xie, “VGGT-Long: Chunk it, Loop it, Align it–Pushing VGGT’s Limits on Kilometer-scale Long RGB Sequences,”arXiv:2507.16443, 2025

    Pith/arXiv arXiv 2025

  26. [26]

    GigaSLAM: Large-Scale Monocular SLAM with Hierarchical Gaussian Splats,

    K. Deng, Y . Zhang, J. Yang, and J. Xie, “GigaSLAM: Large-Scale Monocular SLAM with Hierarchical Gaussian Splats,” inProc. ACM SIGGRAPH Asia Conf., 2025

    2025

  27. [27]

    DeRF: Decomposed Radiance Fields,

    D. Rebain, W. Jiang, S. Yazdani, K. Li, K. M. Yi, and A. Tagliasacchi, “DeRF: Decomposed Radiance Fields,” inProc. IEEE/CVF Conf. on Computer Vision and Pattern Recognition, 2021

    2021

  28. [28]

    Block-NeRF: Scalable Large Scene Neural View Synthesis,

    M. Tancik, V . Casser, X. Yan, S. Pradhan, B. Mildenhall, P. P. Srinivasan, J. T. Barron, and H. Kretzschmar, “Block-NeRF: Scalable Large Scene Neural View Synthesis,” inProc. IEEE/CVF Conf. on Computer Vision and Pattern Recognition, 2022

    2022

  29. [29]

    LiteFlowNet3: Resolving Correspondence Ambiguity for More Accurate Optical Flow Estimation,

    T.-W. Hui and C. C. Loy, “LiteFlowNet3: Resolving Correspondence Ambiguity for More Accurate Optical Flow Estimation,” inProc. European Conf. on Computer Vision, 2020

    2020

  30. [30]

    VIGS-SLAM: Visual Inertial Gaussian Splatting SLAM,

    Z. Zhu, W. Zhang, M. Li, N. Haala, M. Pollefeys, and D. Barath, “VIGS-SLAM: Visual Inertial Gaussian Splatting SLAM,” inProc. European Conf. on Computer Vision, 2026

    2026

  31. [31]

    SEGS-SLAM: Structure-enhanced 3D Gaussian Splatting SLAM with Appearance Embedding,

    T. Wen, Z. Liu, and Y . Fang, “SEGS-SLAM: Structure-enhanced 3D Gaussian Splatting SLAM with Appearance Embedding,” inProc. IEEE/CVF Int. Conf. on Computer Vision, 2025

    2025

  32. [32]

    Vision meets Robotics: The KITTI Dataset,

    A. Geiger, P. Lenz, C. Stiller, and R. Urtasun, “Vision meets Robotics: The KITTI Dataset,”Int. Journal of Robotics Research, vol. 32, no. 11, pp. 1231–1237, Sep. 2013

    2013

  33. [33]

    1 Year, 1000km: The Oxford RobotCar Dataset,

    W. Maddern, G. Pascoe, C. Linegar, and P. Newman, “1 Year, 1000km: The Oxford RobotCar Dataset,”Int. Journal of Robotics Research, vol. 36, no. 1, pp. 3–15, Jan. 2017

    2017

  34. [34]

    The M ´alaga urban dataset: High-rate stereo and LiDAR in a realistic urban scenario,

    J.-L. Blanco-Claraco, F. ´Angel Moreno-Due ˜nas, and J. Gonz ´alez- Jim´enez, “The M ´alaga urban dataset: High-rate stereo and LiDAR in a realistic urban scenario,”Int. Journal of Robotics Research, vol. 33, no. 2, pp. 207–214, Feb. 2014

    2014

This paper was first reviewed by deepseek-v4-flash on August 1, 2026.