Pith. sign in

REVIEW 4 major objections 5 minor 40 references

CHOW-SLAM: Compact Hybrid Representation with Complementary Overlap Window Optimization for RGB-D SLAM

T0 review · 4 major / 5 minor · reviewed 2026-08-04 · deepseek-v4-flash

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

Pith's one-line read CHOW-SLAM claims that a compact parametric–hash scene representation plus a complementary overlap-window keyframe policy lets online neural RGB-D SLAM beat state-of-the-art reconstruction and tracking on Replica, ScanNet, and TUM RGB-D.

desk verdict A competent integration of hybrid representations and keyframe-window scheduling with plausible reconstruction gains, but the tracking claims are confounded by unablated ORB initialization and the closest baselines are missing. read the letter →

arxiv 2608.01914 v1 pith:XZR2UC26 submitted 2026-08-03 cs.CV

Wenxuan Ji , Jin Xiao , Xiaoguang Hu , Jiaqi Shi , Zichong Jia , Baochang Zhang This is my paper
classification cs.CV
keywords NeRF-basedSLAMRGB-Dcompacthybridrepresentationoverlapwindowoptimizationdensereconstructioncameratrackingkeyframeselectionbundleadjustment
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

CHOW-SLAM sets out to resolve a specific tension in neural-radiance-field SLAM: with a limited online budget, a system must build spatial constraints that are compact yet detailed enough for dense reconstruction, and temporal constraints that are persistent enough to prevent drift and forgetting. The paper's proposal is to fix the two sides separately — a scale-aware 'parametric–hash' (P-H) hybrid representation of the scene for space, and a 'complementary overlap-window' strategy for time that mixes recent frames, high-overlap local keyframes, and historically distributed keyframes within a fixed window. It then reports that on Replica the method reaches an average depth error of 0.84 cm, a completion ratio of 96.75%, and a trajectory error of 0.56 cm — the best numbers among the neural SLAM systems compared — and that it also leads the neural baselines on ScanNet and TUM RGB-D. A sympathetic reader would care because the paper is arguing that dense neural mapping and accurate tracking are not in conflict: the right representation and the right frame-selection policy can close much of the tracking gap to classical feature-based SLAM without sacrificing dense, appearance-aware maps.

What carries the argument

Central object: the compact P-H hybrid representation — a parametric branch (coarse explicit 3D grids plus fine orthogonal feature planes) and a hash branch (coarse 2D hash planes plus a fine 3D hash grid) — whose concatenated features feed one small multi-output decoder predicting TSDF, density, and color. A latent geometry code carries TSDF information into the density-color branch, and a symmetric-KL loss aligns the TSDF-derived and density-derived ray-termination distributions so both views of where a ray stops agree. Second mechanism: complementary overlap-window optimization (COWO), which fills a fixed 20-keyframe window from five subsets — recent frames, top-overlap local keyframes, r

What would settle it

Re-run the Replica and ScanNet experiments with the ORB initializer disabled — warm-starting neural tracking from the previous frame's pose — and keep everything else fixed; if trajectory error stays near 0.56 cm and 6.94 cm, the representation and COWO carry the gains, while a rise toward the Co-SLAM/QQ-SLAM baseline levels would implicate ORB. A complementary check: apply the Replica-tuned thresholds (tau_kf = 2.2, delta_BA = 0.2) to ScanNet and TUM without recalibration; sharp degradation would show the published numbers depend on per-dataset tuning.

Watch

Extended reading notes

Core claim

On its own terms, the paper's discovery is that online neural SLAM quality is set by two coupled choices — how the scene is encoded and which historical frames are re-optimized — and that both should be scale-aware and complementary rather than simply larger. Spatially, CHOW-SLAM pairs a parametric branch (coarse 3D grids plus fine planes) with a hash branch (coarse 2D hash planes plus a fine 3D hash grid), decoded by one small MLP that predicts TSDF, density, and color while aligning the TSDF- and density-derived ray-termination distributions via symmetric KL loss. Temporally, it splits a 20-keyframe window into recent, top-overlap, randomized-overlap, and fixed/random historical frames, an

Load-bearing premise

The reported superiority assumes the gains come from the new P-H representation and overlap-window optimization rather than from the ORB-based pose initialization, which is never ablated, and from keyframe-insertion and bundle-adjustment thresholds tuned separately for each dataset.

Editorial extensions

If this is right

  • If the central claim holds, NeRF-based SLAM no longer has to trade compactness against detail: the scale-aware P-H split reports finer surfaces and more complete meshes than hash-only or tri-plane-only encodings at a moderate parameter count (14.72 M).
  • Keyframe selection becomes an accuracy lever in its own right: the paper's ablation shows that removing the full COWO scheme raises the absolute trajectory error on Replica from 0.56 cm to 0.75 cm — a larger effect than any single representation change tested.
  • Classical feature tracking can serve as a lightweight pose initializer rather than a full parallel SLAM pipeline; with it, the paper reports the system narrows the TUM tracking gap to ORB-SLAM2 (1.77 cm vs 1.00 cm average) while keeping dense appearance-aware mapping.
  • The reported combination — 0.84 cm depth error, 96.75% completion ratio, and 12.5 FPS on Replica — indicates the system is within reach of interactive dense-mapping use, not just offline reconstruction.

Reading between the lines

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

  • A direct way to test the tracking contribution of each stage: disable the ORB initialization (Section 3.5) and warm-start neural tracking from the previous frame's pose; the split between the classical initializer and the neural refinement is not measured in the paper.
  • The thresholds tau_kf and delta_BA are tuned per dataset (2.2/1.2/1.3 and 0.2/0.15/0.6 for Replica, ScanNet, TUM), so the method's cross-scene generality is an open question; a self-tuning scheduler based on running loss statistics would make the approach transfer without manual calibration.
  • The complementary-window idea is representation-agnostic: 3D-Gaussian-Splatting SLAM systems, which typically pick keyframes by overlap alone, could adopt the recent/local/historical split to improve long-term consistency.
  • The TSDF-density alignment loss is a generic consistency principle that could strengthen monocular or dynamic-scene neural SLAM, where depth supervision is weaker, without adding networks.
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. CHOW-SLAM is an online dense RGB-D neural SLAM system. It proposes a compact parametric–hash (P-H) hybrid scene representation, a complementary overlap-window optimization (COWO) strategy for keyframe and bundle-adjustment selection, and a two-stage tracking scheme in which ORB-based feature tracking provides an initial pose that is then refined by neural rendering optimization. The paper reports state-of-the-art reconstruction and tracking results on Replica, ScanNet, and TUM RGB-D, with, for example, a Replica average Depth L1 of 0.84 cm, Completion Ratio of 96.75%, and ATE RMSE of 0.56 cm. The source code is released.

Significance. As a systems paper, the work has a clearly described method and the central reconstruction result is empirically plausible: the scale-aware combination of a coarse-grid/fine-plane parametric branch with a coarse-plane/fine-hash branch is a reasonable design, and the complementary window appears to help maintain both local and historical constraints. The level of implementation detail is sufficient for reimplementation, and the code release is a genuine strength. However, the stronger claims in the abstract and contributions are not yet supported by the experiments. The closest prior hybrid systems, HS-SLAM and MHED-SLAM, are absent from the comparison tables; the ORB-based pose initialization is never ablated; several key thresholds are tuned per dataset; and no repeated runs or uncertainty intervals are provided despite the stochastic components in the window selection. These gaps are load-bearing because they directly affect whether the claimed improvements over the state of the art are attributable to the proposed P-H representation and COWO rather than to existing components or to tuning. None of these issues appear to be unfixable within the scope of the manuscript, so I recom

major comments (4)
  1. [§3.5, Tables 2, 4, and 5] The tracking claim is confounded by the ORB-based initialization. Every run, including all ablations in Table 5, uses the ORB pose initialization described in §3.5, and no experiment disables it. Moreover, the abstract's claim that CHOW-SLAM 'outperforms state-of-the-art methods in camera tracking accuracy' is contradicted by the paper's own tables: on Replica, SplaTAM has lower average ATE (0.36 cm vs. 0.56 cm, Table 2), and on TUM, classical ORB-SLAM2 has lower average ATE (1.00 cm vs. 1.77 cm, Table 4). The authors should provide an ORB-off ablation, and ideally an ORB-only neural-refinement-off ablation, to separate the contribution of the proposed representation and COWO from the classical pose prior. Without this, the tracking half of the central claim is not established.
  2. [§4.1.3, Tables 1–4] The two closest prior hybrid systems, HS-SLAM [22] and MHED-SLAM [43], are mentioned in Related Work and are the direct sources of the H-branch architecture and the self-supervised consistency loss (Sec. 3.1 and Sec. 3.6), yet neither appears in any comparison table. A paper claiming to outperform the state of the art in RGB-D neural SLAM cannot omit the immediate predecessors of its own representation. The authors should add these baselines, or at minimum provide a clear and quantitative comparison on the same evaluation protocol, before claiming representational superiority.
  3. [§4.1.4, Tables 1–4] Several load-bearing hyperparameters are tuned per dataset: tau_kf = 2.2/1.2/1.3 and delta_BA = 0.2/0.15/0.6 for Replica/ScanNet/TUM. The paper reports no sensitivity analysis and no cross-dataset transfer test, so the reported gains of COWO and loss-aware BA could reflect dataset-specific tuning rather than the proposed mechanism. A parameter sweep or a transfer test (e.g., using Replica-tuned thresholds on ScanNet and TUM) is needed to support the robustness claim.
  4. [§3.4.1, Tables 5 and 6] The window selection includes randomized subsets (K_lrand and K_hrand in Eq. (9)), and neural optimization is stochastic, but all quantitative results appear to be single runs with no error bars or repeated-seed statistics. Several differences in the ablation table are small (e.g., Depth L1 of 0.87 vs. 0.85 vs. 0.84; ATE of 0.64 vs. 0.67 vs. 0.56), and the paper states that removing COWO components causes 'clear degradation' without statistical support. Please report at least three independent runs per configuration and report mean±std or confidence intervals for the main tables and the ablation.
minor comments (5)
  1. [Eq. (22)] The symmetric KL divergence expression is missing parentheses around the fractions; as written, 'wi log wi + ε / ŵi + ε' is ambiguous. It should be written as log((wi+ε)/(ŵi+ε)) and similarly for the second term.
  2. [§4.2.1, Table 1 caption] The table caption says 'Our method achieves the highest tracking accuracy,' but Table 2 shows SplaTAM with lower average ATE on the same dataset. The text in §4.2.1 correctly qualifies this ('except the 3DGS-based SplaTAM'); the caption and Table 1 note should be made consistent.
  3. [§4.1.4] The exact values of tau_BA are not reported; the text only says it is 'slightly larger' than tau_kf. Since the paper introduces loss-aware BA as a contribution, the concrete threshold values should be listed.
  4. [§3.4.2 and §3.5] The minimum trigger interval of 2 frames for adaptive keyframe insertion and the 10-frame cooldown for BA are stated without justification. A sentence explaining the choice would improve reproducibility and help readers understand the sensitivity of the method.
  5. [§5, Limitations] The limitations paragraph lists dynamic environments and large-scale scenes, but does not mention the absence of an ORB-off ablation or the dependency on classical feature tracking in texture-poor regions, which is more directly tied to the claims of this paper. This could be acknowledged.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: empirical benchmarks provide independent evidence; no prediction reduces by construction to a fitted parameter or self-citation.

full rationale

CHOW-SLAM is an empirical systems paper whose central claims are supported by measurements on external benchmarks (Replica, ScanNet, TUM). The P-H hybrid representation is assembled from components explicitly credited to prior work (HS-SLAM [22], MHED-SLAM [43]), and the SKL consistency loss is likewise inherited from [43]; these are antecedent contributions rather than self-citations, and the benchmarks provide independent falsifiability. The only step resembling fitting is per-dataset tuning of τ_kf and δ_BA, but these are scheduler thresholds, not outputs renamed as predictions; no quantity in Tables 1–5 is the value of an optimized parameter by construction. The absence of an ORB-off ablation and the omission of HS-SLAM/MHED-SLAM from comparisons are attribution and completeness limitations, not circularity: the tracking numbers are empirical measurements, and the ORB initialization is a pipeline component whose contribution is untested, not a parameter that definitionally forces the reported ATE. No load-bearing derivation reduces to its own input, so the circularity score is 0.

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

The system is an empirical combination of known building blocks. There are no invented physical entities, and the central claim rests on external benchmarks. The load-bearing choices are the per-dataset hyperparameters and the domain assumptions that overlap scores and SKL alignment are useful optimization signals.

free parameters (5)
  • Per-dataset keyframe and BA thresholds (tau_kf, tau_BA, delta_BA) = tau_kf: 2.2 (Replica), 1.2 (ScanNet), 1.3 (TUM); delta_BA: 0.2, 0.15, 0.6
    Chosen separately for each dataset and used to decide keyframe insertion and loss-aware BA triggering. They directly affect which frames are optimized and thus influence the reported tracking and mapping numbers.
  • Sliding-window quotas (W, N_r, rho, N_hfix) = W=20, N_r=2, rho=0.4, N_hfix=2
    Hand-chosen decomposition of the optimization window into recent, local, and historical subsets. No sensitivity analysis is provided.
  • P-H representation resolutions and channels = Geometry coarse grid 24 cm and fine planes 6 cm; appearance 24 cm and 3 cm; 32 channels; 16-level hash with 8 plane leve
    Architecture hyperparameters that determine compactness and reconstruction detail. They are fixed without a sweep, so the optimality of this configuration is not established.
  • Loss weights = Mapping: lambda_fs=5, lambda_center=200, lambda_tail=10, lambda_d=0.1, lambda_c=5; tracking: 10, 200, 50, 1, 5; SKL init
    Manual weighting of color, depth, TSDF, and self-supervision terms. These values balance geometry, appearance, and tracking and are not swept.
  • Per-dataset optimization budget = Mapping 10 iterations (Replica, ScanNet) or 20 (TUM); tracking 10 iterations; rays 1024 or 2048; stratified samples N_u
    Different compute budgets across datasets complicate fair comparison and are chosen per dataset.
assumptions (5)
  • domain assumption The TSDF-to-opacity conversion in Eq. (6), with a sigmoid controlled by learnable beta, is a valid surface model.
    Standard in neural SLAM; the sharpness parameter beta is learned during optimization.
  • standard math NeRF-style volume rendering (Eqs. 7-8) with stratified and surface-guided sampling yields unbiased color and depth estimates.
    This is the standard differentiable rendering model introduced by NeRF and used throughout neural SLAM.
  • domain assumption The overlap score in Eq. (12), computed by projecting sampled 3D points into candidate keyframes, is a reliable proxy for the optimization value of a historical frame.
    Used to allocate local and historical keyframe quotas; the paper does not validate that this proxy correlates with actual tracking or mapping gain.
  • domain assumption The symmetric KL alignment in Eq. (22) between TSDF-derived and density-derived ray termination distributions improves both geometry and appearance without hurting tracking.
    Adopted from MHED-SLAM [43] and treated as a given rather than derived or ablated in this paper.
  • domain assumption ORB-based pose initialization (Section 3.5) is reliable in the tested indoor scenes and does not bias the neural refinement step.
    This component is never ablated, and the paper itself lists texture-deficient regions as a limitation.

how reviews work

0 comments
Cite this review

Pith. "Pith review of CHOW-SLAM: Compact Hybrid Representation with Complementary Overlap Window Optimization for RGB-D SLAM." pith.science (2026). https://pith.science/paper/XZR2UC26

@misc{pith2026260801914,
  author       = {Pith},
  title        = {Pith review of: CHOW-SLAM: Compact Hybrid Representation with Complementary Overlap Window Optimization for RGB-D SLAM},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/XZR2UC26}},
  note         = {Machine review of arXiv:2608.01914}
}
read the original abstract

Simultaneous localization and mapping (SLAM) based on Neural Radiance Fields (NeRF) enables dense, continuous scene reconstruction. However, existing systems operating with limited online resources struggle to simultaneously construct two types of constraints, namely, compact yet discriminative spatial constraints derived from scene representations and persistent temporal constraints derived from historical observations. To address this challenge, we propose CHOW-SLAM, a dense RGB-D SLAM framework that explicitly constructs these complementary spatial and temporal constraints. Spatially, we propose a compact parametric-hash (P-H) hybrid representation that organizes components based on planes and grids across scales in P and H branches. A unified multi-output decoder further aligns the ray termination distributions induced by TSDF and density, preserving geometry and appearance under a compact parameter budget. Temporally, we propose a complementary overlap-window strategy to prevent optimization from being dominated by short-term overlap or weakly related historical observations. Within a fixed budget, the strategy retains recent frames, selects high-overlap local frames, and introduces temporally distributed historical keyframes. Loss-aware keyframe insertion and bundle adjustment scheduling further adapt optimization to tracking quality. In addition, ORB-based tracking and geometric pose estimation are used for pose initialization, followed by neural rendering optimization to improve tracking stability. Extensive evaluations on multiple datasets demonstrate that CHOW-SLAM outperforms state-of-the-art methods in both scene reconstruction quality and camera tracking accuracy. The source code is available at https://github.com/jinjidexiaohuoban/CHOW-SLAM.

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

40 extracted references · 3 canonical work pages

  1. [22]

    Z. Gong, F. Tosi, Y . Zhang, S. Mattoccia, M. Poggi, HS-SLAM: Hybrid repre- sentation with structural supervision for improved dense SLAM, in: Proc. IEEE Int. Conf. Robot. Autom., 2025, pp. 8464–8470.https://doi.org/10.1109/ ICRA55743.2025.11127551

    arXiv 2025

  2. [43]

    D. Feng, W. Qin, Z. Shi, W. Chen, Y . Duan, L. Xie, E. Yin, MHED-SLAM: Multi-scale hybrid encoding-based decoupled SLAM, in: Proc. AAAI Conf. Artif. Intell., vol. 40, 2026, pp. 18243–18252.https://doi.org/10.1609/ aaai.v40i22.38887

    2026

  3. [1]

    Wang, K.H

    Y . Wang, K.H. Yu, J.T. Jia, Y . Nie, Enhanced visual SLAM and path planning for autonomous navigation of wheeled mobile robots, J. Vis. Exp. 224 (2025) e68794.https://doi.org/10.3791/68794

    work page doi:10.3791/68794 2025

  4. [2]

    Y . Liu, L. Liu, Y . Zheng, Y . Liu, F. Dang, N. Li, et al., Embodied navi- gation, Sci. China Inf. Sci. 68 (2025) 141101.https://doi.org/10.1007/ s11432-024-4303-8

    2025

  5. [3]

    Campos, R

    C. Campos, R. Elvira, J.J. Gómez Rodríguez, J.M.M. Montiel, J.D. Tardós, ORB-SLAM3: An accurate open-source library for visual, visual-inertial, and multimap SLAM, IEEE Trans. Robot. 37 (6) (2021) 1874–1890.https://doi. org/10.1109/TRO.2021.3075644

    arXiv 2021

  6. [4]

    T. Qin, P. Li, S. Shen, VINS-Mono: A robust and versatile monocular visual- inertial state estimator, IEEE Trans. Robot. 34 (4) (2018) 1004–1020.https: //doi.org/10.1109/TRO.2018.2853729

    arXiv 2018

  7. [5]

    Engel, T

    J. Engel, T. Schöps, D. Cremers, LSD-SLAM: Large-scale direct monocular SLAM, in: Proc. Eur. Conf. Comput. Vis., 2014, pp. 834–849.https://doi. org/10.1007/978-3-319-10605-2_54

    doi:10.1007/978-3-319-10605-2_54 2014

  8. [6]

    Z. Lin, Q. Zhang, Z. Tian, P. Yu, Z. Ye, H. Zhuang, J. Lan, SLAM2: Simultane- ous localization and multimode mapping for indoor dynamic environments, Pat- tern Recognit. 158 (2025) 111054.https://doi.org/10.1016/j.patcog. 2024.111054

    arXiv 2025

Show all 40 references
  1. [7]

    Muñoz-Salinas, R

    R. Muñoz-Salinas, R. Medina-Carnicer, UcoSLAM: Simultaneous localiza- tion and mapping by fusion of keypoints and squared planar markers, Pat- tern Recognit. 101 (2020) 107193.https://doi.org/10.1016/j.patcog. 2019.107193

    2020

  2. [8]

    Z. Teed, J. Deng, DROID-SLAM: Deep visual SLAM for monocular, stereo, and RGB-D cameras, Adv. Neural Inf. Process. Syst. 34 (2021) 16558–16569. 28

    2021

  3. [9]

    Czarnowski, T

    J. Czarnowski, T. Laidlow, R. Clark, A.J. Davison, DeepFactors: Real-time probabilistic dense monocular SLAM, IEEE Robot. Autom. Lett. 5 (2) (2020) 721–728.https://doi.org/10.1109/LRA.2020.2965415

    2020

  4. [10]

    Y . Fan, Q. Zhang, Y . Tang, S. Liu, H. Han, Blitz-SLAM: A semantic SLAM in dynamic environments, Pattern Recognit. 121 (2022) 108225.https://doi. org/10.1016/j.patcog.2021.108225

    2022

  5. [11]

    L. Xia, X. Li, Z. Wang, H. Chen, X. Zhu, L. Fan, DMSAA-SLAM: RGB-D SLAM for dynamic scenes via diffusion self-attention, Pattern Recognit. 179 (2026) 113576.https://doi.org/10.1016/j.patcog.2026.113576

    2026

  6. [12]

    Mildenhall, P.P

    B. Mildenhall, P.P. Srinivasan, M. Tancik, J.T. Barron, R. Ramamoorthi, R. Ng, NeRF: Representing scenes as neural radiance fields for view synthesis, in: Proc. Eur. Conf. Comput. Vis., 2020, pp. 405–421.https://doi.org/10.1007/ 978-3-030-58452-8_24

    2020

  7. [13]

    S. Lai, L. Cui, J. Yin, Fast radiance field reconstruction from sparse inputs, Pat- tern Recognit. 157 (2025) 110863.https://doi.org/10.1016/j.patcog. 2024.110863

    2025

  8. [14]

    Z. Zhu, S. Peng, V . Larsson, W. Xu, H. Bao, Z. Cui, M.R. Oswald, M. Polle- feys, NICE-SLAM: Neural implicit scalable encoding for SLAM, in: Proc. IEEE/CVF Conf. Comput. Vis. Pattern Recognit., 2022, pp. 12776–12786. https://doi.org/10.1109/CVPR52688.2022.01245

    2022

  9. [16]

    Müller, A

    T. Müller, A. Evans, C. Schied, A. Keller, Instant neural graphics primitives with a multiresolution hash encoding, ACM Trans. Graph. 41 (4) (2022) 102. https://doi.org/10.1145/3528223.3530127. 29

    2022

  10. [17]

    Sandström, Y

    E. Sandström, Y . Li, L. Van Gool, M.R. Oswald, Point-SLAM: Dense neural point cloud-based SLAM, in: Proc. IEEE/CVF Int. Conf. Comput. Vis., 2023, pp. 18433–18444.https://doi.org/10.1109/ICCV51070.2023.01690

    2023

  11. [18]

    H. Wang, J. Wang, L. Agapito, Co-SLAM: Joint coordinate and sparse paramet- ric encodings for neural real-time SLAM, in: Proc. IEEE/CVF Conf. Comput. Vis. Pattern Recognit., 2023, pp. 13293–13302.https://doi.org/10.1109/ CVPR52729.2023.01277

    2023

  12. [19]

    G. Li, Q. Chen, Y . Yan, J. Pu, EC-SLAM: Effectively constrained neural RGB-D SLAM with TSDF hash encoding and joint optimization, Pattern Recognit. 170 (2026) 112034.https://doi.org/10.1016/j.patcog.2025.112034

    2026

  13. [20]

    Jiang, J

    S. Jiang, J. Hua, Z. Han, Query quantized neural SLAM, in: Proc. AAAI Conf. Artif. Intell., vol. 39, 2025, pp. 4057–4065.https://doi.org/10.1609/ aaai.v39i4.32425

    2025

  14. [24]

    Ortiz, A

    J. Ortiz, A. Clegg, J. Dong, E. Sucar, D. Novotny, M. Zollhoefer, M. Mukadam, iSDF: Real-time neural signed distance fields for robot perception, in: Proc. Robot. Sci. Syst., 2022.https://doi.org/10.15607/RSS.2022.XVIII. 012. 30

    2022 doi

  15. [25]

    Straub, T

    J. Straub, T. Whelan, L. Ma, Y . Chen, E. Wijmans, S. Green, J.J. Engel, R. Mur-Artal, C. Ren, S. Verma, et al., The Replica dataset: A digital replica of in- door spaces, arXiv preprint arXiv:1906.05797 (2019).https://doi.org/10. 48550/arXiv.1906.05797

    1906 doi arXiv

  16. [26]

    Dai, A.X

    A. Dai, A.X. Chang, M. Savva, M. Halber, T. Funkhouser, M. Nießner, Scan- Net: Richly-annotated 3D reconstructions of indoor scenes, in: Proc. IEEE Conf. Comput. Vis. Pattern Recognit., 2017, pp. 5828–5839.https://doi.org/10. 1109/CVPR.2017.261

    2017

  17. [27]

    Cartillier, G

    V . Cartillier, G. Schindler, I. Essa, SLAIM: Robust dense neural SLAM for on- line tracking and mapping, in: Proc. IEEE/CVF Conf. Comput. Vis. Pattern Recognit. Workshops, 2024, pp. 2862–2871.https://doi.org/10.1109/ CVPRW63382.2024.00292

    2024

  18. [28]

    Mur-Artal, J.D

    R. Mur-Artal, J.D. Tardós, ORB-SLAM2: An open-source SLAM system for monocular, stereo, and RGB-D cameras, IEEE Trans. Robot. 33 (5) (2017) 1255–1262.https://doi.org/10.1109/TRO.2017.2705103

    2017

  19. [29]

    Davison, I.D

    A.J. Davison, I.D. Reid, N.D. Molton, O. Stasse, MonoSLAM: Real-time single camera SLAM, IEEE Trans. Pattern Anal. Mach. Intell. 29 (6) (2007) 1052– 1067.https://doi.org/10.1109/TPAMI.2007.1049

    2007

  20. [30]

    Newcombe, S.J

    R.A. Newcombe, S.J. Lovegrove, A.J. Davison, DTAM: Dense tracking and mapping in real-time, in: Proc. IEEE Int. Conf. Comput. Vis., 2011, pp. 2320– 2327.https://doi.org/10.1109/ICCV.2011.6126513

    2011

  21. [31]

    Whelan, M

    T. Whelan, M. Kaess, H. Johannsson, M. Fallon, J.J. Leonard, J. Mc- Donald, Real-time large-scale dense RGB-D SLAM with volumetric fusion, Int. J. Robot. Res. 34 (4–5) (2015) 598–626.https://doi.org/10.1177/ 0278364914551008

    2015

  22. [32]

    Whelan, S

    T. Whelan, S. Leutenegger, R.F. Salas-Moreno, B. Glocker, A.J. Davison, Elas- ticFusion: Dense SLAM without a pose graph, in: Proc. Robot. Sci. Syst., 2015. https://doi.org/10.15607/RSS.2015.XI.001. 31

    2015 doi

  23. [33]

    Schöps, T

    T. Schöps, T. Sattler, M. Pollefeys, BAD SLAM: Bundle adjusted direct RGB-D SLAM, in: Proc. IEEE/CVF Conf. Comput. Vis. Pattern Recognit., 2019, pp. 134–144.https://doi.org/10.1109/CVPR.2019.00022

    2019

  24. [34]

    X. Yang, H. Li, H. Zhai, Y . Ming, Y . Liu, G. Zhang, V ox-Fusion: Dense tracking and mapping with voxel-based neural implicit representation, in: Proc. IEEE Int. Symp. Mixed Augmented Reality, 2022, pp. 499–507.https://doi.org/10. 1109/ISMAR55827.2022.00066

    2022

  25. [35]

    H. Park, M. Park, G. Nam, J. Kim, LRSLAM: Low-rank representation of signed distance fields in dense visual SLAM system, in: Proc. Eur. Conf. Comput. Vis., 2024, pp. 225–240.https://doi.org/10.1007/978-3-031-72989-8_13

    2024 doi

  26. [36]

    Kerbl, G

    B. Kerbl, G. Kopanas, T. Leimkühler, G. Drettakis, 3D Gaussian splatting for real-time radiance field rendering, ACM Trans. Graph. 42 (4) (2023) 139. https://doi.org/10.1145/3592433

    2023 doi

  27. [38]

    Z. Peng, T. Shao, Y . Liu, J. Zhou, Y . Yang, J. Wang, K. Zhou, RTG-SLAM: Real-time 3D reconstruction at scale using Gaussian splatting, in: Proc. ACM SIGGRAPH Conf. Papers, 2024, pp. 1–11.https://doi.org/10.1145/ 3641519.3657455

    2024

  28. [41]

    C. Yan, D. Qu, D. Xu, B. Zhao, Z. Wang, D. Wang, X. Li, GS-SLAM: Dense visual SLAM with 3D Gaussian splatting, in: Proc. IEEE/CVF Conf. Comput. Vis. Pattern Recognit., 2024, pp. 19595–19604.https://doi.org/10.1109/ CVPR52733.2024.01853

    2024

  29. [42]

    H. Wang, Y . Shou, L. Shen, S. Li, Y . Cao, RGD-SLAM: Robust Gaussian splat- ting SLAM for dynamic environments, Pattern Recognit. 175 (2026) 113071. https://doi.org/10.1016/j.patcog.2026.113071

    2026

  30. [44]

    Hedman, P.P

    P. Hedman, P.P. Srinivasan, B. Mildenhall, J.T. Barron, P. Debevec, Bak- ing neural radiance fields for real-time view synthesis, in: Proc. IEEE/CVF Int. Conf. Comput. Vis., 2021, pp. 5855–5864.https://doi.org/10.1109/ ICCV48922.2021.00582

    2021

  31. [45]

    Chung, Y .-C

    C.-M. Chung, Y .-C. Tseng, Y .-C. Hsu, X.-Q. Shi, Y .-H. Hua, J.-F. Yeh, W.- C. Chen, Y .-T. Chen, W.H. Hsu, Orbeez-SLAM: A real-time monocular vi- sual SLAM with ORB features and NeRF-realized mapping, in: Proc. IEEE Int. Conf. Robot. Autom., 2023, pp. 9400–9406.https://doi.o...

    2023

  32. [46]

    Sturm, N

    J. Sturm, N. Engelhard, F. Endres, W. Burgard, D. Cremers, A benchmark for the evaluation of RGB-D SLAM systems, in: Proc. IEEE/RSJ Int. Conf. Intell. Robots Syst., 2012, pp. 573–580.https://doi.org/10.1109/IROS.2012. 6385773. 33

    2012 doi

Pith tools

Reviewed August 4, 2026 · model on record in the stance chip above.