Pith. sign in

REVIEW 5 major objections 6 minor 41 references

MAC-Ego3D: Multi-Agent Gaussian Consensus for Real-Time Collaborative Ego-Motion and Photorealistic 3D Reconstruction

T0 review · 5 major / 6 minor · reviewed 2026-08-11 · deepseek-v4-flash

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

Pith's one-line read MAC-Ego3D claims real-time, photorealistic multi-agent SLAM by aligning shared 3D Gaussian splat maps across agents, cutting trajectory error tenfold and improving rendering PSNR by 4 to 10 dB.

desk verdict A serious and plausible multi-agent dense SLAM system built on Gaussian splats, but the inter-agent alignment bootstrap is omitted and the evaluation is looser than the headlines suggest. read the letter →

arxiv 2412.09723 v1 pith:NAI7KIUK submitted 2024-12-12 cs.CV

Xiaohao Xu , Feng Xue , Shibo Zhao , Yike Pan , Sebastian Scherer , Xiaonan Huang This is my paper · ORCID
classification cs.CV
keywords multi-agentSLAM3DGaussiansplattingcollaborativemappingego-motionestimationphotorealisticreconstructionRGB-Dconsensusreal-time
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

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

The reading

The paper introduces MAC-Ego3D, a multi-agent SLAM system in which each robot builds and refines its own map as a cloud of 3D Gaussian splats, then periodically aligns and jointly optimizes those clouds with neighboring agents. The authors claim this is the first real-time, full-cycle multi-agent dense SLAM, with pose tracking 15 times faster than the prior state of the art and photorealistic reconstruction quality 4 to 10 dB higher in PSNR. If right, teams of robots or augmented-reality devices could build shared, high-fidelity 3D maps on the fly instead of exchanging sparse point clouds or waiting for offline optimization. The framework unifies local tracking and global multi-agent alignment under one consensus protocol, which is what makes the speed and fidelity gains plausible.

What carries the argument

The central mechanism is Multi-Agent Gaussian Consensus, a four-step protocol of Selection, Association, Alignment, and Optimization that runs in two modes: Intra-Agent, which enforces temporal coherence among neighboring splats within one agent, and Inter-Agent, which aligns overlapping local maps by associating co-visible Gaussian splats. The load-bearing objects are 3D Gaussian splats, each a primitive with mean position, covariance, opacity, and color, which can be rendered differentiably so that visual and geometric losses backpropagate through the entire multi-agent map. This representation is what makes fast inter-agent association and joint optimization possible, because splats can be projected to RGB-D images at low cost.

What would settle it

Run two agents through overlapping trajectories with no known relative pose and no shared map origin, so the first overlap detection must bootstrap alignment from scratch; if the Gaussian association and Eq. (6) alignment diverge or fail to converge, the inter-agent consensus claim collapses.

Watch

Extended reading notes

Core claim

MAC-Ego3D demonstrates that a unified 3D Gaussian splat representation can serve both local ego-motion tracking and global multi-agent map alignment in a single optimization loop. Each agent independently estimates pose and builds a local splat map through Intra-Agent Gaussian Consensus, while parallelized Inter-Agent Gaussian Consensus detects overlaps from compact image embeddings, associates Gaussian splats across maps, aligns them by minimizing a covariance-weighted distance, and jointly optimizes maps and poses via differentiable rendering. On the Multi-agent Replica benchmark the paper reports average ATE of 0.14 cm versus 1.23 cm for the prior neural baseline CP-SLAM, and 40.04 dB versus 29.33 dB PSNR; on the real-world 7-Scenes dataset it reports roughly 4 dB PSNR gains and avoids the baseline's partial tracking failures. The central claim is that consensus over Gaussian primitives, applied both within and across agents, yields higher accuracy, higher fidelity, and higher speed than previous sparse or implicit-neural collaborative SLAM systems.

Load-bearing premise

The system assumes that when two agents detect an overlap by image similarity, they can associate their Gaussian splats and solve the relative transform directly, yet the paper never says how the initial relative pose between agents is obtained before that association.

Editorial extensions

If this is right

  • Multi-robot teams can build a single, globally consistent photorealistic map online instead of merging maps after the fact.
  • Loop closure and drift correction happen by aligning Gaussian splats, so trajectory error drops by roughly an order of magnitude on overlapping indoor sequences.
  • The differentiable splat map makes joint multi-agent optimization practical at 0.02 s per frame for tracking and 0.05 s per frame for mapping, according to the paper's measurements.
  • Rendering quality on real indoor scenes improves by about 4 dB PSNR over the previous dense neural baseline, with depth error roughly halved.
  • The same Gaussian representation supports asynchronous inter-agent communication, since overlap detection only requires exchanging compact 1D image embeddings.

Reading between the lines

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

  • A testable extension would run two agents with no shared initial pose reference and no common map origin, forcing the first overlap detection to bootstrap inter-agent alignment from scratch; the paper does not specify how that bootstrap is obtained.
  • If the speedup survives scaling to more than two agents, Gaussian-splat consensus could replace implicit neural maps as the default representation for collaborative dense SLAM, where per-frame mapping time has historically been the bottleneck.
  • The inter-agent alignment loss resembles a weighted rigid registration of Gaussian means, so the same machinery could transfer directly to loop-closure handling within a single agent without new infrastructure.
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

5 major / 6 minor

Summary. The paper proposes MAC-Ego3D, a multi-agent RGB-D SLAM system in which each agent builds a local 3D Gaussian splat map, tracks its ego-motion, and periodically aligns and jointly optimizes overlapping local maps through Intra-Agent and Inter-Agent Gaussian Consensus. The method is evaluated on the Multi-agent Replica and 7-Scenes datasets against CP-SLAM and several sparse multi-agent SLAM baselines, reporting large improvements in ATE, RGB rendering metrics, depth L1 error, and per-frame runtime. The paper also includes robustness tests under RGB-D corruption, ablation studies of collaborative map and pose optimization, and hyperparameter sensitivity analyses. The abstract claims a new state of the art with a 15x inference speedup, order-of-magnitude trajectory error reductions, and 4-10 dB PSNR gains.

Significance. If the reported results hold, MAC-Ego3D would be a meaningful step for collaborative dense SLAM: it demonstrates that 3D Gaussian splats can serve as a unified representation for per-agent tracking, local mapping, inter-agent alignment, and joint refinement, with rendering-based optimization fast enough for real-time operation. The paper's positive aspects include a clearly formulated consensus framework, experiments on both synthetic and real-world benchmarks, robustness analyses, and hyperparameter sensitivity studies. The main caveats are that the inter-agent alignment procedure, which is the load-bearing component of the collaboration, lacks a stated initialization step, and that several evaluation and presentation gaps (no error bars, training-frame-only reconstruction metrics, unlabeled ablation charts) currently make the headline claims difficult to verify from the manuscript alone.

major comments (5)
  1. [Sec. 3.5, Eq. (6)] Inter-Agent Gaussian Consensus, as written, cannot start: after Eq. (5) detects overlap, the association step in Sec. 3.5.2 assumes Gaussians from different agents can be matched by spatial proximity, but no initial relative pose or common coordinate frame is specified. Eq. (6) therefore minimizes over correspondences that are not defined until a candidate transformation exists. The only nearby implementation detail, the Generalized-ICP distance threshold in Sec. 4.1, is described for point selection during tracking and is not connected to the inter-agent association step. Please specify the bootstrap (e.g., an initial transform from G-ICP, a visual place recognition pose, or a handover pose), and provide a sensitivity test, for example by perturbing the initial relative transform. This is load-bearing because every collaborative gain, including the 0.14 cm ATE and the PSNR improvements, depends on correct inter-agent association.
  2. [Tables 2-4] The quantitative claims rest on single-run point estimates without variance or repeatability information. In particular, Table 4 reports several FAIL entries for CP-SLAM but then averages MAC-Ego3D over all scenes, while CP-SLAM's average excludes failures, so the comparison is not apples-to-apples. For a claim of 'new SOTA' with order-of-magnitude margins, please report standard deviations over multiple runs (or state explicitly that the synthetic protocol is deterministic), and provide a per-scene comparison on the subset of sequences where both methods track successfully. Without this, the headline averages are not yet supported.
  3. [Tables 3 and 4, Sec. 4.2] The dense mapping metrics are computed on the same RGB-D frames used to optimize the map, i.e., on the input observations Z_a_i^1:t, so PSNR, SSIM, and LPIPS primarily measure fitting rather than novel-view generalization. The appendix's novel-view rendering results are qualitative only. The abstract's claim of 'RGB PSNR gains of 4 to 10 dB' therefore overstates generalization unless held-out-view metrics are reported. Please add quantitative novel-view PSNR, SSIM, and LPIPS on held-out frames, or explicitly relabel Tables 3 and 4 as training-frame reconstruction quality.
  4. [Abstract and Table 5] The '15x increase in inference speed' is ambiguous. Table 5 reports per-frame tracking time 0.02 s versus CP-SLAM's 0.30 s (15x) and per-frame mapping time 0.05 s versus 10.10 s (roughly 200x). Please clarify whether the speedup claim refers to tracking only, and report end-to-end wall-clock latency for the complete multi-agent pipeline, including communication, overlap detection, association, and alignment, to substantiate the real-time claim.
  5. [Figs. 8 and 9] The ablation charts contain no numeric labels, so the reader cannot recover the exact values of ATE, PSNR, SSIM, and LPIPS for the with- and without-collaboration conditions. Since these ablations are the primary evidence for the benefit of Inter-Agent Gaussian Consensus, one of the paper's two main contributions, please report the corresponding numbers in a table, ideally with variance across runs.
minor comments (6)
  1. [Sec. 3.5, Eq. (5)] The notation p(I_ai, I_aj | H_ai,aj) = S(v_ai, v_aj) equates a probability with a similarity score; please use a similarity-score notation such as s(v_ai, v_aj) to avoid confusion.
  2. [Eq. (8)] The relative transformation T_i,j in Eq. (8) is never defined; it should be linked to the T_aiaj from Eq. (6) and explicitly constrained to SE(3), and the ambiguity between the indices (i,j) in G and the agent indices a_i, a_j should be resolved.
  3. [Sec. C.3] The construction of the 7-Scenes testing cases, including which sequences are paired and why, is only described in the appendix; this information is needed in the main text or in the supplementary material for reproducibility of Tables 2-4.
  4. [Abstract] The paper states that code 'will be made publicly available' but no code or reproducibility instructions are provided; please finalize the code release or provide a detailed pseudocode-level implementation appendix before publication.
  5. [Figures L-O] The axis labels and legends in the hyperparameter sensitivity figures render as garbled glyphs in the provided PDF; please replace them with readable text labels.
  6. [Sec. 3.2 and Sec. 4.2] There are several typographical and formatting issues, including 'throughInter-Agent' with a missing space in Sec. 3.2, and inconsistent table alignment in Table 4 that makes the FAIL entries and averages hard to parse.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the central claims are empirical, benchmarked against external datasets, and no load-bearing derivation reduces to a fitted parameter or self-citation.

full rationale

MAC-Ego3D's central claims (ATE, PSNR, runtime) are evaluated against external benchmarks (Multi-agent Replica and 7-Scenes), not derived from a theory whose inputs contain those claims. The method sections define a Gaussian map representation, intra-agent tracking by Mahalanobis alignment (Eq. 4), inter-agent association and alignment (Eq. 6), and a photometric/geometric loss (Eq. 7). None of these equations defines its output in terms of the metric being reported, nor is any benchmark number a fitted parameter renamed as a prediction. The RGB-D rendering metrics are computed on the same frames used to optimize the map, which is standard practice for dense SLAM reconstruction fidelity and applies equally to the CP-SLAM baseline; it is not a claim of novel-view prediction. The robustness evaluation follows a noise model from a prior paper that shares authors ([33]), but this is a benchmarking protocol, not load-bearing evidence for the method's novelty. The inter-agent consensus (Sec. 3.5) leaves unspecified how an initial relative pose is obtained before spatial-proximity association, but this is an omitted initialization / completeness gap, not a circular reduction: no equation in the paper defines the association as a function of the output transformation it is used to estimate. No self-citation chain is used to justify the core representation or the SOTA claim. Therefore the paper's derivation is self-contained with respect to circularity; concerns about the unstated bootstrap belong to correctness risk, not circularity.

Assumptions & free parameters 6 free parameters · 5 assumptions · 0 invented entities

The system is an empirical integration of published components. The central claim depends on several hand-set hyperparameters, on the transferability of a pre-trained overlap detector, and on an unstated bootstrap for inter-agent association. No new physical entities are introduced.

free parameters (6)
  • Similarity threshold tau = 0.8
    Used to decide whether two agents share a view; hand-set default from sensitivity analysis in Appendix B.1.
  • G-ICP distance threshold = 0.02
    Used for point selection in Generalized-ICP; set in Section 4.1 Implementation details.
  • Loss weights (lambda_L1, lambda_SSIM, lambda_geom) = 0.8, 0.2, 0.1
    Weights for photometric, SSIM, and geometric consistency terms in Eq. (7); chosen by hand, no sensitivity analysis reported.
  • Global pose weight alpha = 1
    Balances map consistency and pose alignment in Eq. (8).
  • Communication interval Tcomm = 150 (Replica), 60 (7-Scenes)
    Inter-agent synchronization period; dataset-specific defaults with sensitivity analysis in Appendix B.2.
  • Pruning thresholds (tau_o, tau_s, tau_e) = 0.005, 0.25, 10
    Opacity, scale, and elongation pruning thresholds in Appendix C.1; these affect map compactness and rendering quality.
assumptions (5)
  • standard math The 3D Gaussian splatting rendering model (Eq. 2) is an accurate approximation of the scene appearance.
    Taken from Kerbl et al. [13] without modification.
  • domain assumption Observed 3D points are Gaussian-distributed with covariances that justify the Mahalanobis-distance alignment in Eq. (6).
    This is a modeling choice inherited from G-ICP; if point uncertainties are badly misspecified, the alignment weights are wrong.
  • domain assumption The joint optimization over Gaussians and poses in Eqs. (7)-(8) converges to a meaningful consensus rather than a poor local minimum.
    Gradient descent on non-convex photometric/geometric losses is used without convergence guarantees.
  • domain assumption The pre-trained visual place recognition embedding [9] transfers to the Replica and 7-Scenes test environments at threshold tau=0.8.
    Overlap detection depends entirely on this embedding; no fine-tuning or failure analysis is reported.
  • ad hoc to paper An initial relative pose between agents is available or recoverable so that spatial-proximity Gaussian association is correct.
    The inter-agent association step (Sec. 3.5.2) is described only after 'overlaps are detected'; the bootstrap transformation is not specified.

how reviews work

0 comments
Cite this review

Pith. "Pith review of MAC-Ego3D: Multi-Agent Gaussian Consensus for Real-Time Collaborative Ego-Motion and Photorealistic 3D Reconstruction." pith.science (2026). https://pith.science/paper/NAI7KIUK

@misc{pith2026241209723,
  author       = {Pith},
  title        = {Pith review of: MAC-Ego3D: Multi-Agent Gaussian Consensus for Real-Time Collaborative Ego-Motion and Photorealistic 3D Reconstruction},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/NAI7KIUK}},
  note         = {Machine review of arXiv:2412.09723}
}
read the original abstract

Real-time multi-agent collaboration for ego-motion estimation and high-fidelity 3D reconstruction is vital for scalable spatial intelligence. However, traditional methods produce sparse, low-detail maps, while recent dense mapping approaches struggle with high latency. To overcome these challenges, we present MAC-Ego3D, a novel framework for real-time collaborative photorealistic 3D reconstruction via Multi-Agent Gaussian Consensus. MAC-Ego3D enables agents to independently construct, align, and iteratively refine local maps using a unified Gaussian splat representation. Through Intra-Agent Gaussian Consensus, it enforces spatial coherence among neighboring Gaussian splats within an agent. For global alignment, parallelized Inter-Agent Gaussian Consensus, which asynchronously aligns and optimizes local maps by regularizing multi-agent Gaussian splats, seamlessly integrates them into a high-fidelity 3D model. Leveraging Gaussian primitives, MAC-Ego3D supports efficient RGB-D rendering, enabling rapid inter-agent Gaussian association and alignment. MAC-Ego3D bridges local precision and global coherence, delivering higher efficiency, largely reducing localization error, and improving mapping fidelity. It establishes a new SOTA on synthetic and real-world benchmarks, achieving a 15x increase in inference speed, order-of-magnitude reductions in ego-motion estimation error for partial cases, and RGB PSNR gains of 4 to 10 dB. Our code will be made publicly available at https://github.com/Xiaohao-Xu/MAC-Ego3D .

Figures

Figures reproduced from arXiv: 2412.09723 by the authors.

Figure 1
Figure 1. Towards collaborative, real-time, photorealistic 3D reconstruction in multi-agent systems. (a) In MAC-Ego3D, each agent independently captures observations, estimates ego-motion, and constructs a local Gaussian-based 3D map, which is then pe￾riodically aligned with others for collaborative optimization. (b) Multi-Agent Gaussian Consensus, with intra- and inter-agent Gaus￾sian selection, association, alignment, and o… view at source ↗
Figure 2
Figure 2. Pipeline overview of MAC-Ego3D. MAC-Ego3D leverages parallel Intra-Agent Gaussian Consensus and periodic Inter-Agent Gaussian Consensus to enable real-time pose tracking and photorealistic 3D reconstruction using a shared 3D Gaussian map representation. vancements in representation techniques [10, 24, 35, 39, 40]. Some recent methods also separate learning-based pose es￾timation from dense mapping [29, 37]. Recently… view at source ↗
Figure 3
Figure 3. Qualitative RGB image rendering quality comparison between multi-agent SLAM models with dense reconstruction capability, i.e., CP-SLAM and our MAC-Ego3D, on Multi-agent Replica (Left) and 7-Scenes (Right) datasets [PITH_FULL_IMAGE:figures/full_fig_p006_3.png] view at source ↗
Figures from the paper (4 more)
Figure 4
Figure 4. Figure 4: Qualitative depth rendering comparison on 7-Scenes. tracking errors across all cases and agents, significantly out￾performing classical multi-agent sparse mapping methods and the dense CP-SLAM. It improves average trajectory es￾timation accuracy by nearly an order of m…
Figure 5
Figure 5. Figure 5: Multi-agent trajectory estimation results on 7-Scenes [PITH_FULL_IMAGE:figures/full_fig_p007_5.png]
Figure 8
Figure 8. Figure 8: Ablation study on collaborative 3D map optimization. w/ w/o 0.10 0.15 0.20 ATE [cm] ( ) w/ w/o 35.0 37.5 40.0 42.5 PSNR ( ) w/ w/o 0.97 0.98 SSIM ( ) w/ w/o 0.04 0.06 0.08 LPIPS ( ) (a) Multi-agent Replica w/ w/o 0 10 20 30 ATE [cm] ( ) w/ w/o 15 20 25 PSNR ( ) w/ w/o …
Figure 9
Figure 9. Figure 9: Ablation study on collaborative pose optimization. struction. It leverages Multi-Agent Gaussian Consensus to align and optimize 3D Gaussian splatting primitives both within and across agents, resulting in dense and coherent scene representations. Experimental results d…

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

41 extracted references · 34 canonical work pages

  1. [1]

    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. 1, 2, 5, 7

    work page 2021

  2. [2]

    A survey on aerial swarm robotics

    Soon-Jo Chung, Aditya Avinash Paranjape, Philip Dames, Shaojie Shen, and Vijay Kumar. A survey on aerial swarm robotics. IEEE Transactions on Robotics , 34(4):837–855,

    work page

  3. [3]

    Bundlefusion: Real-time glob- ally consistent 3d reconstruction using on-the-fly surface rein- tegration

    Angela Dai, Matthias Nießner, Michael Zollhöfer, Shahram Izadi, and Christian Theobalt. Bundlefusion: Real-time glob- ally consistent 3d reconstruction using on-the-fly surface rein- tegration. ACM Trans. Graph., 36(4), 2017. 2

    work page 2017

  4. [4]

    Real-time rgb-d camera relocalization

    Ben Glocker, Shahram Izadi, Jamie Shotton, and Antonio Criminisi. Real-time rgb-d camera relocalization. In Interna- tional Symposium on Mixed and Augmented Reality (ISMAR). IEEE, 2013. 5, 9

    work page 2013

  5. [5]

    Rgbd gs-icp slam

    Seongbo Ha, Jiung Yeon, and Hyeonwoo Yu. Rgbd gs-icp slam. arXiv preprint arXiv:2403.12550, 2024. 1, 3

    arXiv 2024

  6. [6]

    Rgbd gs-icp slam

    Seongbo Ha, Jiung Yeon, and Hyeonwoo Yu. Rgbd gs-icp slam. In European Conference on Computer Vision, pages 180–197. Springer, 2025. 3, 11

    work page 2025

  7. [7]

    Fast depth densifi- cation for occlusion-aware augmented reality

    Aleksander Holynski and Johannes Kopf. Fast depth densifi- cation for occlusion-aware augmented reality. ACM Transac- tions on Graphics (ToG), 37(6):1–11, 2018. 1

    work page 2018

  8. [8]

    Cp-slam: Collaborative neural point-based slam system

    Jiarui Hu, Mao Mao, Hujun Bao, Guofeng Zhang, and Zhaopeng Cui. Cp-slam: Collaborative neural point-based slam system. Advances in Neural Information Processing Systems, 36, 2024. 1, 2, 5, 6, 7, 9, 11

    work page 2024

Show all 41 references
  1. [9]

    Optimal transport ag- gregation for visual place recognition

    Sergio Izquierdo and Javier Civera. Optimal transport ag- gregation for visual place recognition. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2024. 5

    2024

  2. [10]

    Eslam: Efficient dense slam system based on hy- brid representation of signed distance fields

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

    2023

  3. [11]

    Cvi- slam—collaborative visual-inertial slam

    Marco Karrer, Patrik Schmuck, and Margarita Chli. Cvi- slam—collaborative visual-inertial slam. IEEE Robotics and Automation Letters, 3(4):2762–2769, 2018. 1, 2

    2018

  4. [12]

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

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

    2024

  5. [13]

    3d gaussian splatting for real-time radiance field rendering

    Bernhard Kerbl, Georgios Kopanas, Thomas Leimkühler, and George Drettakis. 3d gaussian splatting for real-time radiance field rendering. ACM Transactions on Graphics, 42(4), 2023. 2, 3

    2023

  6. [14]

    Swarm-slam: Sparse decentralized collaborative simultaneous localization and mapping framework for multi-robot systems

    Pierre-Yves Lajoie and Giovanni Beltrame. Swarm-slam: Sparse decentralized collaborative simultaneous localization and mapping framework for multi-robot systems. IEEE Robotics and Automation Letters , 9(1):475–482, 2023. 1, 2, 5 12

    2023

  7. [15]

    Towards collaborative simultaneous localiza- tion and mapping: a survey of the current research landscape

    Pierre-Yves Lajoie, Benjamin Ramtoula, Fang Wu, and Gio- vanni Beltrame. Towards collaborative simultaneous localiza- tion and mapping: a survey of the current research landscape. arXiv preprint arXiv:2108.08325, 2021. 1

    2021 arXiv

  8. [16]

    Loopy-slam: Dense neural slam with loop closures

    Lorenzo Liso, Erik Sandström, Vladimir Yugay, Luc Van Gool, and Martin R Oswald. Loopy-slam: Dense neural slam with loop closures. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , pages 20363–20373, 2024. 3

    2024

  9. [17]

    Gaussian splatting slam

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

    2024

  10. [18]

    Nerf: Representing scenes as neural radiance fields for view syn- thesis

    Ben Mildenhall, Pratul P Srinivasan, Matthew Tancik, Jonathan T Barron, Ravi Ramamoorthi, and Ren Ng. Nerf: Representing scenes as neural radiance fields for view syn- thesis. Communications of the ACM , 65(1):99–106, 2021. 2

    2021

  11. [19]

    Raúl Mur-Artal and Juan D. Tardós. Orb-slam2: An open- source slam system for monocular, stereo, and rgb-d cameras. IEEE Transactions on Robotics, 33(5):1255–1262, 2017. 2

    2017

  12. [20]

    Orb-slam: a versatile and accurate monocular slam system

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

  13. [21]

    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. In 10th IEEE International Symposium on Mixed a...

    2011

  14. [22]

    Compressed 3d gaussian splatting for accelerated novel view synthesis

    Simon Niedermayr, Josef Stumpfegger, and Rüdiger West- ermann. Compressed 3d gaussian splatting for accelerated novel view synthesis. In Proceedings of the IEEE/CVF Con- ference on Computer Vision and Pattern Recognition, pages 10349–10358, 2024. 8

    2024

  15. [23]

    Nerf- slam: Real-time dense monocular slam with neural radiance fields

    Antoni Rosinol, John J Leonard, and Luca Carlone. Nerf- slam: Real-time dense monocular slam with neural radiance fields. arXiv preprint arXiv:2210.13641, 2022. 2

    2022 arXiv

  16. [24]

    Os- wald

    Erik Sandström, Yue Li, Luc Van Gool, and Martin R. Os- wald. Point-slam: Dense neural point cloud-based slam. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), 2023. 3

    2023

  17. [25]

    Ccm-slam: Robust and efficient centralized collaborative monocular simultaneous localization and mapping for robotic teams

    Patrik Schmuck and Margarita Chli. Ccm-slam: Robust and efficient centralized collaborative monocular simultaneous localization and mapping for robotic teams. Journal of Field Robotics, 36(4):763–781, 2019. 1, 2, 5

    2019

  18. [26]

    Generalized-icp

    Aleksandr Segal, Dirk Haehnel, and Sebastian Thrun. Generalized-icp. In Robotics: science and systems , page

  19. [27]

    The replica dataset: A digital replica of indoor spaces

    Julian Straub, Thomas Whelan, Lingni Ma, Yufan Chen, Erik Wijmans, Simon Green, Jakob J Engel, Raul Mur-Artal, Carl Ren, Shobhit Verma, et al. The replica dataset: A digital replica of indoor spaces. arXiv preprint arXiv:1906.05797,

    1906 arXiv

  20. [28]

    imap: Implicit mapping and positioning in real-time

    Edgar Sucar, Shikun Liu, Joseph Ortiz, and Andrew J Davison. imap: Implicit mapping and positioning in real-time. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 6229–6238, 2021. 2

    2021

  21. [29]

    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. Advances in neural information processing systems, 34:16558–16569, 2021. 3

    2021

  22. [30]

    Kimera- multi: Robust, distributed, dense metric-semantic slam for multi-robot systems

    Yulun Tian, Yun Chang, Fernando Herrera Arias, Carlos Nieto-Granda, Jonathan P How, and Luca Carlone. Kimera- multi: Robust, distributed, dense metric-semantic slam for multi-robot systems. IEEE Transactions on Robotics, 38(4),

  23. [31]

    How, and Luca Carlone

    Yulun Tian, Yun Chang, Fernando Herrera Arias, Carlos Nieto-Granda, Jonathan P. How, and Luca Carlone. Kimera- multi: Robust, distributed, dense metric-semantic slam for multi-robot systems. IEEE Transactions on Robotics, 38(4): 2022–2038, 2022. 1, 2

    2022

  24. [32]

    How, and Luca Carlone

    Yulun Tian, Yun Chang, Long Quang, Arthur Schang, Carlos Nieto-Granda, Jonathan P. How, and Luca Carlone. Resilient and distributed multi-robot visual slam: Datasets, experi- ments, and lessons learned. In 2023 IEEE/RSJ International Conference on Intelligent Robots and System...

    2023

  25. [33]

    Customizable perturbation synthesis for robust slam benchmarking

    Xiaohao Xu, Tianyi Zhang, Sibo Wang, Xiang Li, Yongqi Chen, Ye Li, Bhiksha Raj, Matthew Johnson-Roberson, and Xiaonan Huang. Customizable perturbation synthesis for robust slam benchmarking. arXiv preprint arXiv:2402.08125,

    arXiv

  26. [34]

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

    Xingrui Yang, Hai Li, Hongjia Zhai, Yuhang Ming, Yuqian Liu, and Guofeng Zhang. V ox-fusion: Dense tracking and mapping with voxel-based neural implicit representation. In 2022 IEEE International Symposium on Mixed and Aug- mented Reality (ISMAR) , pages 499–507. IEEE, 2022. 5, 7

    2022

  27. [35]

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

    Xingrui Yang, Hai Li, Hongjia Zhai, Yuhang Ming, Yuqian Liu, and Guofeng Zhang. V ox-fusion: Dense tracking and mapping with voxel-based neural implicit representation. In IEEE International Symposium on Mixed and Augmented Reality, ISMAR 2022, Singapore, October 17-21, 2022, ...

    2022

  28. [36]

    The unreasonable effectiveness of deep features as a perceptual metric

    Richard Zhang, Phillip Isola, Alexei A Efros, Eli Shechtman, and Oliver Wang. The unreasonable effectiveness of deep features as a perceptual metric. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 586–595, 2018. 12

    2018

  29. [37]

    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. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), 2023. 3

    2023

  30. [38]

    Loopsplat: Loop closure by registering 3d gaus- sian splats

    Liyuan Zhu, Yue Li, Erik Sandström, Konrad Schindler, and Iro Armeni. Loopsplat: Loop closure by registering 3d gaus- sian splats. arXiv preprint arXiv:2408.10154, 2024. 3

    2024 arXiv

  31. [39]

    Nice-slam: Neural implicit scalable encoding for slam

    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. In Proceedings of the IEEE/CVF Conference on Computer Vi- sion and Pattern Recognition, pages 12786–12796, 2...

    2022

  32. [40]

    Coslam: Collaborative visual slam in dynamic environments

    Danping Zou and Ping Tan. Coslam: Collaborative visual slam in dynamic environments. IEEE transactions on pattern analysis and machine intelligence, 35(2):354–366, 2012. 3 14 MAC-Ego3D CP-SLAM Ground-truthMAC-Ego3D CP-SLAM Ground-truth Figure A. Qualitative RGB image rendering...

    2012

  33. [435]

    Seattle, W A, 2009. 5

    2009

Pith tools

Reviewed August 11, 2026 · model on record in the stance chip above.