Pith. sign in

REVIEW 3 major objections 7 minor 41 references

Stronger structured-light depth plus a depth-first SLAM design is enough for practical, robust tracking and reconstruction.

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 ↗

T0 review · grok-4.5

2026-07-31 13:12 UTC pith:NNF5I3HE

load-bearing objection Solid systems paper: better in-loop SL depth plus a deliberately simple depth-first stack; synthetic evidence is clean, real claims lean on T265 as a failure oracle. the 3 major comments →

arxiv 2607.24495 v1 pith:NNF5I3HE submitted 2026-07-27 cs.CV

NSL-SLAM: High-Fidelity Neural Structured-Light Depth for Practical SLAM and Reconstruction

Jiaheng Li , Binsheng Zhang , Xinhai Chang , Wenzheng Chen This is my paper
classification cs.CV
keywords structured lightdepth estimationSLAM3D reconstructionGaussian splattingneural stereo decodingRGB-D
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.

Structured-light cameras already put dense metric depth into millions of devices, but SLAM systems have not fully used that signal. This paper argues that two coordinated moves change the picture. First, feed frozen monocular depth features into the iterative stereo decoding loop itself—not as a cleanup step after decoding—so the depth maps become sharper and more reliable (about 35% lower RMSE than the prior neural structured-light decoder on their Replica-SL benchmark). Second, treat that dense depth as the main tracking signal: register frames with generalized ICP against an online Gaussian map, bring in sparse image matches only when geometry is ambiguous, and run a light bundle-adjustment backend for drift. The claim is mutual reinforcement: better depth makes a simple pipeline work, and a depth-centric pipeline carries the depth gains into the final map. On shared-depth Replica-SL tests the system leads in tracking and gains about 1.6 F-score points in reconstruction; on eight real D435 scenes it is the only compared method that never catastrophically diverges from a reference tracker, at about 21 FPS online.

Core claim

Once neural structured-light depth is made strong enough by injecting monocular priors into iterative decoding, a deliberately simple depth-centric SLAM stack—dense geometric registration as primary signal, sparse visual anchors only for degenerate cases, online Gaussians, and lightweight backend correction—delivers practical robust tracking and higher-quality reconstruction without heavy general RGB-D machinery.

What carries the argument

Visual-guided GICP with adaptive anchor weighting: dense structured-light depth drives generalized ICP against an online 3D Gaussian map (which also supplies anisotropic covariances), while sparse back-projected image matches are scaled by the relative strength of the geometric and anchor Hessians so depth stays primary and vision only rescues weak geometry.

Load-bearing premise

On real data, large drift away from a consumer reference tracker is treated as proof of catastrophic failure, and staying close to that tracker is treated as evidence of success, even though the paper says the reference is not ground truth.

What would settle it

Run the same eight real sequences (or a new set) with a metrically accurate external ground-truth trajectory; if NSL-SLAM then shows large absolute trajectory error or reconstruction error while still hugging the reference tracker—or if a baseline that diverged from the reference is actually closer to ground truth—the real-world stability claim fails.

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

If this is right

  • Commodity structured-light devices can support online SLAM at interactive rates without treating depth as a noisy secondary cue.
  • Shared-depth protocol comparisons become the fair way to isolate SLAM design once neural SL depth is fixed.
  • Injecting foundation monocular features inside the stereo update loop, rather than after it, is a reusable recipe for active depth sensors.
  • Final maps can be built by reprojecting dense per-frame depth along a lightly corrected trajectory instead of relying on the Gaussian map as the sole geometric output.
  • When depth is dense and metric, sparse vision and light BA are residual stabilizers, not the main engine.

Where Pith is reading between the lines

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

  • The same synthetic-only decoder transferring qualitatively to other structured-light sensors suggests a path to multi-vendor SLAM without per-device retraining.
  • If monocular priors keep improving, further depth RMSE cuts may shrink the need for even the sparse-anchor branch on many indoor scenes.
  • Failure modes where planes and weak texture hit geometry and vision together still imply a remaining niche for inertial or multi-view cues the paper leaves open.
  • A public real SL benchmark with true trajectory and mesh GT would turn the current reference-gap metric into a calibrated absolute ranking.

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 / 7 minor

Summary. The paper presents NSL-SLAM, a SLAM system built around neural structured-light (SL) depth. The depth contribution modifies the NSL decoder (RAFT-Stereo-style iterative disparity refinement) by injecting frozen Depth Anything v2 features into the context pathway at every update step, rather than using a foundation model as post-hoc refinement; on the authors' Replica-SL benchmark this reduces depth RMSE by 35% relative to NSL. The system contribution is a depth-centric pipeline: GICP against an online 3D Gaussian map as the primary tracking signal, sparse LightGlue-derived 3D–3D visual anchors with an adaptive weight for geometrically degenerate frames, a lightweight GTSAM keyframe backend, and "soft" feedback of backend corrections through the Gaussian map. On Replica-SL under a shared-depth protocol (all baselines receive the same predicted depth), NSL-SLAM achieves the best or near-best tracking and reconstruction; on 8 real D435 sequences it is the only method with no sequence exceeding a 0.3 m deviation threshold from a T265 reference trajectory; the system runs at 20.9 FPS on an RTX 4090. Ablations attribute an order-of-magnitude ATE reduction to the depth improvement and further gains to anchors, BA, and soft feedback.

Significance. If the results hold, the paper makes a useful and well-evidenced systems point: once neural structured-light decoding is strong enough, a deliberately simple depth-centric SLAM pipeline — dense GICP as the primary signal, sparse visual anchors only as backup, online Gaussians plus lightweight BA — is competitive with much heavier learned systems. The paper ships several practices the field should reward: a shared-depth SLAM protocol that isolates downstream design from upstream depth quality; first-frame-aligned ATE instead of post-hoc best-fit alignment; a clean component ablation (Table 4) showing depth quality dominates (1.55→0.15 m ATE) and each add-on helps; a one-factor hyperparameter sensitivity sweep; and an unusually candid Limitations section. The depth-side contribution (frozen Depth Anything v2 features injected into the iterative RAFT-Stereo decoding loop rather than applied post-hoc) is a modest but sensible architectural delta over NSL and FoundationStereo, with a clear 35% RMSE reduction on the downstream-relevant benchmark. The main soft spots are the real-world evaluation anchored to a T265 reference rather than ground truth, and the fact that the in-domain depth表

major comments (3)
  1. [§4.4, Table 3; Abstract] The real-world headline claims — 'only method that avoids catastrophic failure on all 8 sequences' and '43.3% lower trajectory deviation' — rest entirely on the T265 trajectory as a de facto oracle, evaluated via first-frame-aligned translation gap with a 0.3 m failure threshold. The paper itself states the T265 'is not ground truth: agreement with it does not prove correctness.' This disclosure is commendable but does not close the evidential gap: if T265 drifts or biases in exactly the selected hard conditions (stairs, treadmill, repeated structure), a baseline marked 'failed' could be closer to the true trajectory, and NSL-SLAM's low gaps demonstrate consistency with a visual-inertial reference rather than absolute correctness. Two concrete remedies, either of which would suffice: (a) add at least one independent anchor on a subset of the real sequences — e.g., return-to-start drift o
  2. [§4.2, Table 1; §B (Eq. 10)] Contribution (1) and the abstract claim a depth estimator that reduces RMSE by 35% over NSL, but Table 1 shows this holds only on Replica-SL (0.069→0.045 m). On the in-domain NSL-Test benchmark, NSL-SLAM is actually worse on RMSE (0.079 vs 0.075) and AbsRel (0.028 vs 0.025), better only on MAE and δ1.05. The text in §4.2 acknowledges this honestly and argues Replica-SL is the relevant operating point, which is a defensible position — but then the headline claim should be scoped accordingly, and the argument would be strengthened by evidence that the gain is not specific to the authors' own Replica-SL image-formation model (Eq. 10, with author-chosen α, β, γ, η and Retinex-based reflectance). A depth-only evaluation on one external synthetic SL rendering (e.g., a different pattern/baseline/attenuation model), or at minimum a sensitivity sweep of the Eq. 10 parameters showing the NSL-vs-ou
  3. [§3.3, Eqs. (4)–(5); Table 4] The robustness story — dense GICP primary, sparse anchors 'for geometrically degenerate cases' — is load-bearing for the system claim, but Eq. (5) as written is a scale-normalization rule, not a degeneracy gate. α = clip(tr(H_geo)/tr(H_anc), 10⁻³, 10³) makes the anchor term's Hessian trace proportional to the geometric term's; when geometry becomes weak (small tr(H_geo), e.g., a large plane), α shrinks and the anchor term is suppressed in aggregate rather than amplified. The mechanism can still fix GICP's null directions if off-plane anchors exist, but that is a different mechanism from 'anchors take over when geometry is ambiguous,' and the 6-decade clip range [10⁻³, 10³] functions as a free hyperparameter. Please (i) state precisely what α is intended to do, with a short derivation; (ii) show per-frame or per-sequence α values on a degenerate sequence (e.g., stairs/office0) to demonstr
minor comments (7)
  1. [§4.1 Baselines; Table 2] Shared-depth protocol in Table 2 is a good controlled design, but please state explicitly whether all baselines were run with authors' code and default hyperparameters, and whether any per-scene tuning was permitted — several baselines (e.g., ViPE office1 ATE 0.1549) degrade sharply under the shared depth, and the fairness of the comparison depends on this detail.
  2. [§4.1 Metrics] The first-frame-aligned ATE is the right protocol for online operation, but it should be defined precisely (which frame's pose anchors the alignment, whether scale is fixed by metric depth) and a citation or formula given; readers will otherwise assume Sim(3) Umeyama alignment.
  3. [Abstract; §4.1] Typographical: repeated missing spaces around numbers in the abstract and intro ('by35%on', 'by1.6points', '43.3%lower'); also 'NSL-Test is the synthetic benchmark...' runs directly into 'Replica-SL' in the Datasets paragraph.
  4. [Table 2] Table 2 mixes bold, yellow, and orange highlighting for top-3 values; in grayscale print the yellow/orange distinction is illegible. Consider bold/underline/italic instead.
  5. [§4.5, Table 4] The 'hard feedback' row (0.0265) is worse than no feedback (0.0188), which is a nice result for the soft-feedback design, but the mechanism (temporary position-LR increase for 5 steps, §A.3) deserves one sentence of intuition in the main text — why abrupt pose writes corrupt the Gaussian registration target while gradual absorption does not.
  6. [§C.1, Fig. 9] Cross-sensor transfer (D415, Orbbec Gemini 335L) is shown only qualitatively for one capture each; either temper the generalization language in §3.2 or add quantitative error against sensor depth on planar/target objects.
  7. [§4.6] Runtime claim (20.9 FPS) should note that CPU GICP and frontend–backend synchronization dominate the non-depth latency, and whether the system remains real-time on hardware below an RTX 4090 — 'practical' in the title invites this question.

Circularity Check

0 steps flagged

Empirical systems paper with no derivation-by-construction circularity; claims are benchmarked, not defined into the premises.

full rationale

NSL-SLAM is an engineering/systems paper: a stronger SL depth decoder (frozen DAv2 features injected into RAFT-Stereo-style iterative decoding) plus a depth-centric SLAM stack (GICP primary, sparse anchors, online Gaussians, lightweight BA). There is no first-principles derivation chain in which a claimed prediction reduces to its inputs by definition or by a fitted parameter renamed as a forecast. Depth gains are measured against NSL, LibSGM, and LingBot-Depth on NSL-Test and Replica-SL with ground-truth depth; full-system tracking/reconstruction uses first-frame-aligned ATE and F-score under an explicit shared-depth protocol that gives all SLAM baselines the same depth, isolating pipeline design. Ablations (LibSGM→NSL depth→anchors→BA→soft feedback) are incremental empirical measurements, not tautologies. Self-citation to NSL is ordinary prior-work scaffolding, not a load-bearing uniqueness theorem that forbids alternatives. Dependence on authors’ Replica-SL render and on T265 as a real-world reference is an evaluation-validity concern, not circularity of derivation: the paper does not define success as matching a quantity it fitted, nor does any equation force the reported RMSE/ATE/F-score. Score 0 is the honest finding.

Axiom & Free-Parameter Ledger

5 free parameters · 5 axioms · 3 invented entities

Claims rest on standard multi-view geometry, off-the-shelf networks and registration, and engineering choices (adaptive anchor weight, soft pose feedback) rather than new physical entities. Load-bearing domain assumptions are that calibrated single-shot SL triangulation remains the metric backbone, frozen monocular features help active IR matching under distribution shift, and T265 divergence indicates catastrophic SLAM failure. Free parameters are the usual SLAM/training knobs; none is presented as a fundamental constant.

free parameters (5)
  • adaptive anchor weight α clip range = [1e-3, 1e3]
    α = clip(tr(H_geo)/tr(H_anc), α_min, α_max) with α_min=1e-3, α_max=1e3 sets how much sparse vision can override dense depth; central robustness claim depends on this schedule.
  • depth network training hyperparameters = AdamW 6e-5, 200k steps, 16 iters
    LR 6e-5, 200k steps, batch 32, 16 update iters on NSL_train; depth quality (and thus SLAM) depends on this fit to synthetic SL data.
  • GICP/keyframe/backend thresholds = defaults in Supp. A.2–A.4
    Overlap 0.8, mapping every 10 frames, backend min interval 5, motion 0.08 m / 20°, window 20, match counts, Gaussian LRs and 5-step feedback LR boost are hand-set; sensitivity table is limited to two scenes.
  • real-sequence failure threshold = 0.3 m
    Average gap > 0.3 m to T265 marks failure; the 0/8 vs baselines headline depends on this cutoff.
  • Replica-SL image-formation parameters = β auto-calibrated per dataset; baseline 8 cm
    Synthetic SL rendering uses α, β, γ, η, 8 cm baseline, Retinex reflectance; downstream depth/SLAM numbers live on this simulated sensor model.
axioms (5)
  • domain assumption Metric depth from calibrated disparity z = fB/d is the correct geometric backbone for tracking and reconstruction.
    Sec. 3.1 makes triangulation the first determinant of SLAM geometry; the whole depth-centric design assumes this model holds for the sensor pair.
  • domain assumption Frozen monocular DAv2 features provide useful global context on projected-pattern IR without breaking metric stereo grounding.
    Sec. 3.2 chooses DAv2 as most stable on SL/IR among tested models; generalization claims rest on this prior transferring from RGB pretraining.
  • domain assumption Dense GICP against an online Gaussian map is a sufficient primary tracker when per-frame SL depth is strong.
    Sec. 3.3–3.4 follow GS-ICP-style registration; sparse vision and BA are only residual correctors.
  • ad hoc to paper Large first-frame-aligned translation gap to T265 indicates catastrophic tracking failure on real sequences.
    Sec. 4.1 and Table 3 define the real benchmark this way while admitting T265 is not GT.
  • standard math Standard SE(3) registration, factor-graph BA, and 3D Gaussian optimization mathematics hold as used (GICP covariances, GTSAM factors, etc.).
    Eqs. (2)–(6) invoke standard residuals and robust norms without new theorems.
invented entities (3)
  • Replica-SL benchmark no independent evidence
    purpose: Provide synthetic RGB+IR pattern+GT depth+GT poses for joint depth and full-system SLAM evaluation.
    Author-constructed rendering of Replica with a custom projector image-formation model (Supp. B); central quantitative SLAM claims use this set.
  • D435 real 8-scene benchmark with T265 reference no independent evidence
    purpose: Stress-test catastrophic failure modes on real structured-light video where public RGB-D sets lack IR pattern views.
    Collected by authors; labels success via T265 proximity rather than external GT poses/maps.
  • NSL-SLAM depth-centric pipeline (visual-guided GICP + soft Gaussian feedback) no independent evidence
    purpose: System composition that keeps dense SL depth primary while handling degenerate cases and long-range drift.
    Named integration of known modules; not a new physical entity, but a postulated system organization whose value is only evidenced inside this paper’s benches.

pith-pipeline@v1.2.0-grok45-kimik3 · 21145 in / 4262 out tokens · 74865 ms · 2026-07-31T13:12:41.144701+00:00 · methodology

0 comments
read the original abstract

Structured-light (SL) cameras power depth sensing in millions of devices, and recent neural SL decoding methods have substantially improved their depth quality. SLAM systems can benefit greatly from such strong depth sensing, where reliable geometry enables stable tracking and faithful reconstruction. In this work, we present NSL-SLAM, a practical SLAM system tailored for high-fidelity structured-light depth. We first strengthen SL depth sensing: inspired by the neural structured-light (NSL) method, we further incorporate strong monocular depth priors into the SL stereo decoding, reducing depth RMSE by 35% on Replica-SL compared to NSL. We then build a depth-centric SLAM pipeline with this stronger depth: because structured-light geometry is dense and metrically accurate, we keep it as the primary tracking signal, and add only sparse visual correspondences for geometrically degenerate cases and lightweight bundle adjustment for long-range drift. Our depth estimator and SLAM design reinforce each other: stronger depth makes a simple SLAM pipeline effective, and the depth-centric pipeline ensures this advantage transfers to downstream reconstruction. Experimentally, on the synthetic Replica-SL benchmark, NSL-SLAM achieves the best tracking accuracy and improves reconstruction F-score by 1.6 points over the SOTA baseline under a shared-depth protocol. On a real benchmark of 8 challenging scenes, it is the only method that avoids catastrophic failure on all sequences while achieving 43.3% lower trajectory deviation than selected baselines. The SLAM system runs online at 20.9 FPS, demonstrating that stronger structured-light depth and depth-centric system design together enable practical, robust SLAM.

Figures

Figures reproduced from arXiv: 2607.24495 by Binsheng Zhang, Jiaheng Li, Wenzheng Chen, Xinhai Chang.

Figure 1
Figure 1. Figure 1: Our system, NSL-SLAM, captures synchronized RGB and single-shot structured-light video, predicts high-quality [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Pipeline of NSL-SLAM. Structured-light frames are decoded into high-quality depth maps, which drive a depth [PITH_FULL_IMAGE:figures/full_fig_p004_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: Depth estimation in NSL-SLAM. Stereo matching [PITH_FULL_IMAGE:figures/full_fig_p004_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: Qualitative depth comparison on D435 real captures. Depth values are shown in meters. [PITH_FULL_IMAGE:figures/full_fig_p007_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: Real-data reconstruction comparison between a commodity baseline formed by D435 sensor depth with the T265 [PITH_FULL_IMAGE:figures/full_fig_p008_5.png] view at source ↗
Figure 6
Figure 6. Figure 6: Runtime breakdown on an RTX 4090 at 848 × 480. TensorRT depth runs at 27.5 FPS and the full online SLAM system runs at 20.9 FPS. A.4 Hyperparameter Sensitivity The default settings are: 𝛼 by Eq. 5 with clip [10−3 , 103 ], backend window 20, keyframe every 5 frames with 0.08 m/20◦ thresholds, feedback 5 steps, and position LR 1.6 × 10−6 normally / 5 × 10−3 in feedback [PITH_FULL_IMAGE:figures/full_fig_p011… view at source ↗
Figure 7
Figure 7. Figure 7: Additional real-scene depth examples on four D435 sequences. The top row of each scene shows the RGB input, and [PITH_FULL_IMAGE:figures/full_fig_p012_7.png] view at source ↗
Figure 8
Figure 8. Figure 8: Real-data reconstruction comparison between a commodity baseline formed by D435 sensor depth with the T265 [PITH_FULL_IMAGE:figures/full_fig_p013_8.png] view at source ↗
Figure 9
Figure 9. Figure 9: Cross-sensor qualitative transfer from the same synthetic-only NSL_train model to one capture per sensor from the [PITH_FULL_IMAGE:figures/full_fig_p013_9.png] 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

41 extracted references · 19 linked inside Pith

  1. [1]

    Robust single-shot structured light 3D imaging via neural feature decoding

    Jiaheng Li, Qiyu Dai, Lihan Li, Praneeth Chakravarthula, He Sun, Baoquan Chen, and Wenzheng Chen. Robust single-shot structured light 3D imaging via neural feature decoding. InSIGGRAPH Asia 2025 Conference Papers, 2025. arXiv:2512.14028

    arXiv 2025

  2. [2]

    Depth anything V2

    Lihe Yang, Bingyi Kang, Zilong Huang, Zhen Zhao, Xiaogang Xu, Jiashi Feng, and Hengshuang Zhao. Depth anything V2. InAdvances in Neural Information Processing Systems (NeurIPS), volume 37, 2024. arXiv:2406.09414

    Pith/arXiv arXiv 2024

  3. [3]

    Pattern codification strategies in structured light systems.Pattern Recognition, 37(4):827–849, 2004

    Joaquim Salvi, Jordi Pagès, and Joan Batlle. Pattern codification strategies in structured light systems.Pattern Recognition, 37(4):827–849, 2004

    2004

  4. [4]

    Structured-light 3D surface imaging: A tutorial.Advances in Optics and Photonics, 3(2):128–160, 2011

    Jason Geng. Structured-light 3D surface imaging: A tutorial.Advances in Optics and Photonics, 3(2):128–160, 2011

    2011

  5. [5]

    Stereo processing by semiglobal matching and mutual information.IEEE Transactions on Pattern Analysis and Machine Intelligence, 30(2):328–341, 2008

    Heiko Hirschmüller. Stereo processing by semiglobal matching and mutual information.IEEE Transactions on Pattern Analysis and Machine Intelligence, 30(2):328–341, 2008

    2008

  6. [6]

    HyperDepth: Learning depth from structured light without matching

    Sean Ryan Fanello, Christoph Rhemann, Vladimir Tankovich, Adarsh Kowdle, Sergio Orts Escolano, David Kim, and Shahram Izadi. HyperDepth: Learning depth from structured light without matching. In2016 IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 5441–5450, 2016

    2016

  7. [7]

    UltraStereo: Efficient learning-based matching for active stereo systems

    Sean Ryan Fanello, Julien Valentin, Christoph Rhemann, Adarsh Kowdle, Vladimir Tankovich, Philip Davidson, and Shahram Izadi. UltraStereo: Efficient learning-based matching for active stereo systems. InProceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 2691–2700, 2017

    2017

  8. [8]

    ActiveStereoNet: End-to-end self-supervised learning for active stereo systems

    Yinda Zhang, Sameh Khamis, Christoph Rhemann, Julien Valentin, Adarsh Kowdle, Vladimir Tankovich, Michael Schoenberg, Shahram Izadi, Thomas Funkhouser, and Sean Fanello. ActiveStereoNet: End-to-end self-supervised learning for active stereo systems. InEuropean Conference on Computer Vision (ECCV), pages 802–819, 2018. arXiv:1807.06009

    Pith/arXiv arXiv 2018

  9. [9]

    Depth estimation by combining binocular stereo and monocular structured-light

    Yuhua Xu, Xiaoli Yang, Yushan Yu, Wei Jia, Zhaobi Chu, and Yulan Guo. Depth estimation by combining binocular stereo and monocular structured-light. In 2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 1736–1745, 2022

    2022

  10. [10]

    Polka Lines: Learning structured illumination and reconstruction for Active Stereo.CoRR, abs/2011.13117, 2020

    Seung-Hwan Baek and Felix Heide. Polka Lines: Learning structured illumination and reconstruction for Active Stereo.CoRR, abs/2011.13117, 2020

    Pith/arXiv arXiv 2011

  11. [11]

    Masked Depth Modeling for spatial perception, 2026

    Bin Tan, Changjiang Sun, Xiage Qin, Hanat Adai, Zelin Fu, Tianxiang Zhou, Han Zhang, Yinghao Xu, Xing Zhu, Yujun Shen, and Nan Xue. Masked Depth Modeling for spatial perception, 2026

    2026

  12. [12]

    Newcombe, Shahram Izadi, Otmar Hilliges, David Molyneaux, David Kim, Andrew J

    Richard A. Newcombe, Shahram Izadi, Otmar Hilliges, David Molyneaux, David Kim, Andrew J. Davison, Pushmeet Kohli, Jamie Shotton, Steve Hodges, and Andrew W. Fitzgibbon. KinectFusion: Real-time dense surface mapping and tracking. InIEEE International Symposium on Mixed and Augmented Reality (ISMAR), pages 127–136, 2011

    2011

  13. [13]

    Salas-Moreno, Ben Glocker, Andrew J

    Thomas Whelan, Renato F. Salas-Moreno, Ben Glocker, Andrew J. Davison, and Stefan Leutenegger. ElasticFusion: Real-time dense SLAM and light source estimation.The International Journal of Robotics Research, 35(14):1697–1716, 2016

    2016

  14. [14]

    BundleFusion: Real-time globally consistent 3D reconstruction using on-the-fly surface re-integration.ACM Transactions on Graphics, 36(3), 2017

    Angela Dai, Matthias Nießner, Michael Zollöfer, Shahram Izadi, and Christian Theobalt. BundleFusion: Real-time globally consistent 3D reconstruction using on-the-fly surface re-integration.ACM Transactions on Graphics, 36(3), 2017

    2017

  15. [15]

    ReFusion: 3D reconstruction in dynamic environments for RGB-D cameras exploiting residuals

    Emanuele Palazzolo, Jens Behley, Philipp Lottes, Philippe Giguère, and Cyrill Stachniss. ReFusion: 3D reconstruction in dynamic environments for RGB-D cameras exploiting residuals. InIEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), pages 7855–7862, 2019

    2019

  16. [16]

    Gómez Rodríguez, José M

    Carlos Campos, Richard Elvira, Juan J. Gómez Rodríguez, José M. M. Montiel, and Juan D. Tardós. ORB-SLAM3: An accurate open-source library for visual, visual- inertial, and multimap SLAM.IEEE Transactions on Robotics, 37(6):1874–1890, 2021

    2021

  17. [17]

    BAD SLAM: Bundle adjusted direct RGB-D SLAM

    Thomas Schöps, Torsten Sattler, and Marc Pollefeys. BAD SLAM: Bundle adjusted direct RGB-D SLAM. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 134–144, 2019

    2019

  18. [18]

    DROID-SLAM: Deep visual SLAM for monocular, stereo, and RGB-D cameras

    Zachary Teed and Jia Deng. DROID-SLAM: Deep visual SLAM for monocular, stereo, and RGB-D cameras. InAdvances in Neural Information Processing Systems (NeurIPS), pages 16558–16569, 2021. arXiv:2108.10869

    Pith/arXiv arXiv 2021

  19. [19]

    MegaSaM: Accurate, fast, and robust structure and motion from casual dynamic videos

    Zhengqi Li, Richard Tucker, Forrester Cole, Qianqian Wang, Linyi Jin, Vickie Ye, Angjoo Kanazawa, Aleksander Holynski, and Noah Snavely. MegaSaM: Accurate, fast, and robust structure and motion from casual dynamic videos. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 10486–10496, 2025. arXiv:2412.04463

    Pith/arXiv arXiv 2025

  20. [20]

    ViPE: Video pose engine for 3d geometric perception, 2025

    Jiahui Huang, Qunjie Zhou, Hesam Rabeti, Aleksandr Korovko, Huan Ling, Xuanchi Ren, Tianchang Shen, Jun Gao, Dmitry Slepichev, Chen-Hsuan Lin, Jiawei Ren, Kevin Xie, Joydeep Biswas, Laura Leal-Taixe, and Sanja Fidler. ViPE: Video pose engine for 3d geometric perception, 2025

    2025

  21. [21]

    Edgar Sucar, Shikun Liu, Joseph Ortiz, and Andrew J. Davison. iMAP: Implicit mapping and positioning in real-time. InProceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 6229–6238, 2021. arXiv:2103.12352

    Pith/arXiv arXiv 2021

  22. [22]

    Oswald, and Marc Pollefeys

    Zihan Zhu, Songyou Peng, Viktor Larsson, Weiwei Xu, Hujun Bao, Zhaopeng Cui, Martin R. Oswald, and Marc Pollefeys. NICE-SLAM: Neural implicit scalable encoding for SLAM. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 12786–12796, 2022. arXiv:2112.12130

    Pith/arXiv arXiv 2022

  23. [23]

    Vox-Fusion: Dense tracking and mapping with voxel-based neural im- plicit representation

    Xingrui Yang, Hai Li, Hongjia Zhai, Yuhang Ming, Yuqian Liu, and Guofeng Zhang. Vox-Fusion: Dense tracking and mapping with voxel-based neural im- plicit representation. InIEEE International Symposium on Mixed and Augmented Reality (ISMAR), pages 499–507, 2022. arXiv:2210.15858

    Pith/arXiv arXiv 2022

  24. [24]

    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. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 13293–13302, 2023. arXiv:2304.14377

    Pith/arXiv arXiv 2023

  25. [25]

    ESLAM: Effi- cient dense SLAM system based on hybrid representation of signed distance fields

    Mohammad Mahdi Johari, Camilla Carta, and François Fleuret. ESLAM: Effi- cient dense SLAM system based on hybrid representation of signed distance fields. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 17408–17419, 2023. arXiv:2211.11704

    Pith/arXiv arXiv 2023

  26. [26]

    Erik Sandström, Yue Li, Luc Van Gool, and Martin R. Oswald. Point-SLAM: Dense neural point cloud-based SLAM. InProceedings of the IEEE/CVF International Con- ference on Computer Vision (ICCV), pages 18387–18398, 2023. arXiv:2304.04278

    Pith/arXiv arXiv 2023

  27. [27]

    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. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 21357– 21366, 2024. arXiv:2312.02126

    Pith/arXiv arXiv 2024

  28. [28]

    Kelly, and Andrew J

    Hidenobu Matsuki, Riku Murai, Paul H.J. Kelly, and Andrew J. Davison. Gaussian splatting SLAM. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 18039–18048, 2024. arXiv:2312.06741

    Pith/arXiv arXiv 2024

  29. [29]

    Photo-SLAM: Real- time simultaneous localization and photorealistic 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 photorealistic mapping for monocular, stereo, and RGB-D cameras. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 21584–21593, 2024

    2024

  30. [30]

    RTG-SLAM: Real-time 3D reconstruction at scale using Gaussian splatting

    Zhexi Peng, Tianjia Shao, Yong Liu, Jingke Zhou, Yin Yang, Jingdong Wang, and Kun Zhou. RTG-SLAM: Real-time 3D reconstruction at scale using Gaussian splatting. InACM SIGGRAPH Conference Proceedings, 2024

    2024

  31. [31]

    CaRtGS: Computational alignment for real-time Gaussian splatting SLAM

    Dapeng Feng, Zhiqiang Chen, Yizhen Yin, Shipeng Zhong, Yuhua Qi, and Hongbo Chen. CaRtGS: Computational alignment for real-time Gaussian splatting SLAM. IEEE Robotics and Automation Letters, 2025. arXiv:2410.00486

    Pith/arXiv arXiv 2025

  32. [32]

    GO-SLAM: Global optimization for consistent 3D instant reconstruction

    Youmin Zhang, Fabio Tosi, Stefano Mattoccia, and Matteo Poggi. GO-SLAM: Global optimization for consistent 3D instant reconstruction. InProceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 3727–3737,

  33. [33]

    Oswald, and Federico Tombari

    Erik Sandström, Keisuke Tateno, Michael Oechsle, Michael Niemeyer, Luc Van Gool, Martin R. Oswald, and Federico Tombari. Splat-SLAM: Globally optimized 9 Li et al. RGB-only SLAM with 3D gaussians.arXiv preprint arXiv:2405.16544, 2024

    Pith/arXiv arXiv 2024

  34. [34]

    LoopSplat: Loop closure by registering 3D gaussian splats

    Liyuan Zhu, Yue Li, Erik Sandström, Shengyu Huang, Konrad Schindler, and Iro Armeni. LoopSplat: Loop closure by registering 3D gaussian splats. In International Conference on 3D Vision (3DV), 2025. arXiv:2408.10154

    Pith/arXiv arXiv 2025

  35. [35]

    RemixFusion: Residual-based mixed repre- sentation for large-scale online RGB-D reconstruction.ACM Transactions on Graphics, 2025

    Yuqing Lan, Chenyang Zhu, Shuaifeng Zhi, Jiazhao Zhang, Zhoufeng Wang, Renjiao Yi, Yijie Wang, and Kai Xu. RemixFusion: Residual-based mixed repre- sentation for large-scale online RGB-D reconstruction.ACM Transactions on Graphics, 2025. arXiv:2507.17594

    arXiv 2025

  36. [36]

    RGBD GS-ICP SLAM

    Seongbo Ha, Jiung Yeon, and Hyeonwoo Yu. RGBD GS-ICP SLAM. InEuropean Conference on Computer Vision (ECCV), pages 180–197, 2024. arXiv:2403.12550

    Pith/arXiv arXiv 2024

  37. [37]

    RAFT-Stereo: Multilevel recurrent field transforms for stereo matching

    Lahav Lipson, Zachary Teed, and Jia Deng. RAFT-Stereo: Multilevel recurrent field transforms for stereo matching. InInternational Conference on 3D Vision (3DV), pages 218–227, 2021. arXiv:2109.07547

    Pith/arXiv arXiv 2021

  38. [38]

    FoundationStereo: Zero-shot stereo matching

    Bowen Wen, Matthew Trepte, Joseph Aribido, Jan Kautz, Orazio Gallo, and Stan Birchfield. FoundationStereo: Zero-shot stereo matching. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 5249–5260, 2025

    2025

  39. [39]

    Generalized-ICP

    Aleksandr Segal, Dirk Haehnel, and Sebastian Thrun. Generalized-ICP. In Robotics: Science and Systems (RSS), 2009

    2009

  40. [40]

    Factor graphs and GTSAM: A hands-on introduction

    Frank Dellaert. Factor graphs and GTSAM: A hands-on introduction. Technical Report GT-RIM-CP&R-2012-002, Georgia Institute of Technology, 2012

    2012

  41. [41]

    Engel, Raúl Mur-Artal, Carl Ren, Shobhit Verber, et al

    Julian Straub, Thomas Whelan, Lingni Ma, Yufan Chen, Erik Wijmans, Simon Green, Jakob J. Engel, Raúl Mur-Artal, Carl Ren, Shobhit Verber, et al. The Replica dataset: A digital replica of indoor spaces.arXiv preprint arXiv:1906.05797, 2019. 10 NSL-SLAM: High-Fidelity Neural Structured-Light Depth for Practical SLAM and Reconstruction Supplementary Material...

    Pith/arXiv arXiv 1906