Pith. sign in

REVIEW 4 major objections 6 minor 56 references

MAGiSt3R claims that multiple agents can build a globally consistent 3D map from uncalibrated monocular RGB video at about 10 FPS by merging learned point-map submaps, and that this outperforms existing feed-forward reconstruction and track

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 · deepseek-v4-flash

2026-08-01 23:49 UTC pith:TPXKDGZU

load-bearing objection A genuine first in multi-agent feed-forward 3R reconstruction, with a real learned merging module; but the reconstruction claim on real video is not quantitatively backed—only tracking is. the 4 major comments →

arxiv 2607.15211 v1 pith:TPXKDGZU submitted 2026-07-16 cs.CV

MAGiSt3R: Multi-Agent Feed-forward 3D Reconstruction from Monocular RGB Videos

Ziren Gong , Xiaohan Li , Fabio Tosi , Ninghui Xu , Stefano Mattoccia , Jianfei Cai , Matteo Poggi This is my paper
classification cs.CV
keywords multi-agent 3D reconstructionfeed-forward 3R modelsmonocular RGB SLAMpoint map mergingpose graph optimizationcamera trackingsubmap fusionloop closure
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

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

The paper tries to establish that feed-forward 3D reconstruction—turning RGB frames into point maps in one pass, without per-scene optimization—can be extended from single cameras to teams of agents. It introduces MAGiSt3R, built on a feed-forward 3R backbone, and a new learned module, MAGMA, that merges local submaps both within one agent over time and across agents when their trajectories overlap. A pose-graph optimization step then corrects drift in the merged camera trajectories. On the ReplicaMultiagent and AriaMultiagent benchmarks, the authors report that MAGiSt3R achieves higher reconstruction accuracy and lower trajectory error than other feed-forward approaches, at nearly 10 FPS. The reason to care is that multi-agent 3D mapping could run online on ordinary RGB cameras without depth sensors, calibration, or per-scene training.

Core claim

The central claim is that one learned module, MAGMA, can serve as the alignment layer for both temporal and multi-agent merging of feed-forward point maps. MAGMA takes a local submap and a retrieved reference set from the global map, fuses geometry tokens with appearance tokens, and regresses both the registered point map and refined camera poses in a single forward pass. Trained on scenes split into two synthetic trajectories and supervised by registration, pose, and geometry-consistency losses, it outperforms geometric aligners such as RANSAC+ICP and SL(4) optimization. Combined with pose-graph optimization, this yields a multi-agent RGB-only reconstruction system that the paper reports as

What carries the argument

The load-bearing component is the Multi-Agent Global Map Aggregation (MAGMA) transformer. It encodes reference and registering point maps into geometry tokens, combines them with appearance tokens derived from visual and spatial tokens through an attention module, and feeds the integrated features to two decoders that predict refined and registered point maps and camera poses. An attention mask lets the registering pose tokens consult reference pose tokens only, forcing alignment into the global frame. Around MAGMA, the system uses a feed-forward 3R backbone for per-submap point maps and a loop-detection step based on global descriptor correlation, followed by pose-graph optimization over SE

Load-bearing premise

The learned merger is trained only on synthetic two-agent trajectories made by splitting single videos from ScanNet, ScanNet++, and Aria Synthetic, so the whole accuracy claim rests on the assumption that this simulation matches the overlap and viewpoint patterns of real multi-agent deployments.

What would settle it

Run MAGiSt3R on a two-agent sequence where the same textured region is observed under strongly different illumination or from very similar viewpoints with no novel visual content, and measure the global point-map alignment error; if merging accuracy collapses to the level of RANSAC+ICP, the claim that appearance tokens resolve geometry ambiguity is falsified.

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

If this is right

  • Multi-agent dense mapping becomes possible with uncalibrated RGB cameras only, eliminating the need for depth sensors and known camera intrinsics in collaborative reconstruction.
  • At roughly 10 FPS, the pipeline is fast enough for near-real-time use, a step beyond RGB-D multi-agent systems that run at only a few frames per second.
  • Existing single-agent feed-forward systems can be lifted to multi-agent operation by swapping their alignment step for MAGMA-like learned merging, as the ablations show it beats geometric alignment baselines.
  • The low-overlap custom demo, with roughly 15% overlap and moving people, suggests the framework can tolerate loosely coordinated agents, though the paper presents that result qualitatively.

Where Pith is reading between the lines

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

  • A natural stress test would be to train MAGMA on simulated trajectories with deliberately varied overlap ratios and measure merging accuracy as overlap shrinks; this would reveal whether the current first/last-frame split is the limiting factor.
  • Because MAGMA consumes point maps plus appearance tokens rather than raw images, the same module could plausibly merge maps from heterogeneous agents, such as different cameras or one RGB and one RGB-D agent, an extension the paper does not test.
  • The loop-detection threshold is fixed at 0.6; making it adaptive to scene appearance would be a testable extension that could matter for larger teams or repetitive indoor environments.

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

4 major / 6 minor

Summary. The manuscript introduces MAGiSt3R, a multi-agent feed-forward dense 3D reconstruction system for monocular RGB videos. Each agent processes short clips with VGGT to produce local point maps, poses, and visual descriptors; a learned module (MAGMA) merges submaps both within an agent and across agents after loop detection, and pose-graph optimization refines global poses. The paper claims state-of-the-art reconstruction and camera tracking accuracy against feed-forward baselines on ReplicaMultiagent and AriaMultiagent, at roughly 10 FPS.

Significance. If the empirical claims hold, this is a plausible first instantiation of feed-forward multi-agent RGB 3D reconstruction, and the component-wise ablations give credible evidence that the learned MAGMA merging module is better than RANSAC+ICP, SL(4), and L2W alternatives. The system does not require camera calibration or per-scene optimization, which is a meaningful advance for collaborative mapping. The main caveats are that reconstruction quality is quantitatively demonstrated only on a synthetic benchmark, and that the training distribution for inter-agent merging is narrow relative to real deployment conditions.

major comments (4)
  1. [Abstract and §4.2 (Tables 6–7)] The claim of 'superior reconstruction and camera tracking accuracy' on synthetic and real-world datasets is not fully supported. Reconstruction accuracy (Acc/Comp) is reported only on ReplicaMultiagent, which is synthetic; the real AriaMultiagent benchmark is used exclusively for ATE tracking because, as stated, it has no ground-truth reconstructions. The abstract and contributions should be rephrased to distinguish synthetic reconstruction results from real-world tracking results, or real-world reconstruction evidence should be supplied. This is load-bearing because the central contribution is an empirical performance claim.
  2. [§4 Dataset and Fig. 5] MAGMA is trained on simulated two-agent trajectories obtained by starting from the first and last frames of each ScanNet/ScanNet++/Aria Synthetic scene and sampling every 2 frames. This yields a specific high-overlap central meeting pattern. The only low-overlap (~15%) evidence is a self-collected sequence presented qualitatively with no baseline. Since the claimed benefit over classical merging is most relevant under low/partial overlap, the paper needs either quantitative low-overlap experiments or a clear statement that generalization to arbitrary overlap regimes is not established. As written, the learned-merging advantage could be tied to the training overlap distribution.
  3. [Tables 1–7] All error metrics are single-point estimates. There are no error bars, repeated runs, or significance tests, and some gaps are small (e.g., Table 4, Agent 1, Apart 0: MAGiSt3R 2.54 vs MASt3R 2.52). The statement that MAGiSt3R 'consistently outperforms' all baselines is therefore stronger than the evidence. Please provide variance estimates or a statistical test, or moderate the claims, especially for the headline ReplicaMultiagent and AriaMultiagent comparisons.
  4. [§3.2, Eq. (9)] The registration loss is described as normalized by the average Euclidean distance and the text mentions scale factors z and \hat z, but the displayed formula contains no normalization and no z terms. The mask M, the confidence C_g, and the '·' operation are also undefined. Because this loss is central to training MAGMA, the inconsistency blocks exact reproduction. Please correct the equation and define all quantities.
minor comments (6)
  1. [§3.2, Eq. (10)] 'where P_g = [q,b,f]' should likely refer to \hat P_g or otherwise distinguish predicted from ground-truth parameters. Please clarify the notation for the quaternion, translation, and field-of-view components.
  2. [§3.1] The phrase 'multiple submaps, e.g., n' appears to be a typo; it should read 'i.e., n' or similar.
  3. [Table 5] The MA-MASt3R-SLAM row has unusual formatting ('- /4.43*') and the footnote explains the average excluding Apart 2. Consider presenting the two averages in separate columns to avoid ambiguity.
  4. [Table 8] Runtime numbers for CP-SLAM and MAGiC-SLAM are taken from original papers while others are measured by the authors. A sentence specifying hardware and evaluation conditions for the measured values would make the comparison fairer.
  5. [Fig. 5] The caption lists '*' with two meanings ('contains moving subjects' and possibly the trajectory markers). Clarify the legend and state whether the low-overlap sequence has any quantitative evaluation or is purely qualitative.
  6. [§6 Limitations] 'Our current framework, as well as the baselines, is not meant for agents facing very different conditions' should be rephrased for clarity. Also, the limitation that only >3 agents are not tested is a plan; please note that the real-world evidence is limited to 3-agent Aria sequences.

Circularity Check

0 steps flagged

No significant circularity; the core claims rest on held-out empirical evaluation rather than fitted or self-referential reductions.

full rationale

MAGiSt3R's central claim is an empirical performance comparison: MAGMA is trained on ScanNet, ScanNet++, and Aria Synthetic with simulated two-agent trajectories, while reconstruction and tracking are evaluated on held-out ReplicaMultiagent and AriaMultiagent. The losses used to train MAGMA (Eqs. 9-12) are standard confidence-aware registration, pose, and geometry-consistency losses, and no test-set constants are fitted. The pose graph optimization (Eq. 13) is a classical refinement applied to MAGMA outputs, not a component that forces the reported accuracy. The self-citations (Ov3R [12], DINO-SLAM [13], HS-SLAM [14], stereo GS-SLAM [25], and the NeRF/3DGS SLAM survey [43]) appear only in related-work context and are not used as load-bearing premises for the claimed first-multi-agent-feed-forward status or for the quantitative results. The paper honestly notes that AriaMultiagent has no ground-truth reconstructions and that the low-overlap custom sequence is only qualitatively evaluated; these are coverage limitations, not circularity. No derivation step in the paper reduces by construction to its own inputs.

Axiom & Free-Parameter Ledger

6 free parameters · 4 axioms · 1 invented entities

The central claim depends on a frozen VGGT backbone, a pretrained SALAD descriptor, and a learned MAGMA module trained on simulated multi-agent data. The free parameters are hyperparameters that are manually set but not fitted to the test benchmarks. The main epistemic risk is the generalization of the simulation-to-real step.

free parameters (6)
  • τconf = 0.25
    Confidence threshold for filtering VGGT point maps; chosen by hand in Implementation Details (Sec. 4).
  • τloop = 0.6
    Correlation threshold for loop detection; chosen by hand and directly affects when inter-agent merges happen.
  • N = 10
    Number of top-correlated views used as reference set in MAGMA; manual choice.
  • β = 1
    Regularization weight in confidence-aware registration loss, following DUSt3R.
  • λreg, λpose, λgeo = 1, 1, 0.8
    Loss weighting terms in Eq. 12; manually set.
  • m and stride = 10 and 5
    Number of frames per submap and sampling stride; manual choices.
axioms (4)
  • domain assumption VGGT provides sufficiently accurate local point maps, depths, and camera parameters from RGB inputs.
    The entire pipeline depends on the quality of the frozen VGGT backbone (Sec. 3.1, Eq. 1). If VGGT's local predictions are unreliable, MAGMA cannot correct them.
  • domain assumption DINOv2 SALAD global descriptors provide reliable visual place recognition for loop detection across agents.
    Loop detection in Sec. 3.3 relies on a correlation threshold over SALAD descriptors. Poor loop detection would break inter-agent merging.
  • domain assumption Simulated multi-agent trajectories from single-scene datasets (ScanNet, ScanNet++, Aria Synthetic) generalize to real multi-agent settings.
    The training protocol in Sec. 4 (Dataset) creates trajectories by sampling from first and last frames; real multi-agent overlap and viewpoint patterns may differ.
  • domain assumption Ground-truth alignment for ATE evaluation is unbiased and does not hide substantial drift.
    ATE RMSE is computed after aligning estimated and ground-truth trajectories, which is standard practice but can be sensitive to the alignment method.
invented entities (1)
  • MAGMA (Multi-Agent Global Map Aggregation) independent evidence
    purpose: Learned model to merge submaps at both intra-agent and inter-agent levels, producing a unified global point map and refined camera poses.
    MAGMA is evaluated on held-out ReplicaMultiagent and AriaMultiagent datasets, giving empirical evidence outside the training set. It is not a hypothetical entity but a concrete learned module.

pith-pipeline@v1.3.0-alltime-deepseek · 15815 in / 9110 out tokens · 75779 ms · 2026-08-01T23:49:14.185325+00:00 · methodology

0 comments
read the original abstract

This paper presents MAGiSt3R, a multi-agent 3D reconstruction framework performing reconstruction and camera tracking for monocular RGB videos at almost 10 FPS. MAGiSt3R relies on a feed-forward model from the 3R family to process RGB videos and regress local point maps, and on a merging model, MAGMA, that combines local maps at both intra-agent and inter-agent levels to obtain the final global point map. Furthermore, MAGiSt3R performs pose graph optimization to mitigate cumulative camera drift occurring along the feed-forward pipeline. We evaluate MAGiSt3R on both synthetic and real-world datasets, demonstrating its superior reconstruction and camera tracking accuracy compared to state-of-the-art approaches.

Figures

Figures reproduced from arXiv: 2607.15211 by Fabio Tosi, Jianfei Cai, Matteo Poggi, Ninghui Xu, Stefano Mattoccia, Xiaohan Li, Ziren Gong.

Figure 1
Figure 1. Figure 1: MAGiSt3R – multi-agent 3D reconstruction in action. [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Overview of MAGiSt3R. Given RGB sequences from multiple agents, our method simultaneously predicts local point maps and camera poses. Then, the MAGMA model merges submaps at the intra-agent level. On the server side, when loops are detected between agents, the same MAGMA module fuses the local maps of multiple agents into a global map. Finally, pose graph optimization further mitigates cumulative drift. qu… view at source ↗
Figure 3
Figure 3. Figure 3: Overview of MAGMA model. Given a global submap, a keyframe set, and a local submap, the MAGMA module merges the local submap into a global map with poses in a unified coordinate system. First, we use retrieval to obtain the most relevant frames to form the inputs, reference set and registering set. Next, we encode their point maps and aggregate these geometric features with visual tokens, spatial tokens, a… view at source ↗
Figure 4
Figure 4. Figure 4: Qualitative results – Dense point maps on ReplicaMultiagent. [PITH_FULL_IMAGE:figures/full_fig_p012_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: Experiments on a self-collected scene with limited overlap. [PITH_FULL_IMAGE:figures/full_fig_p015_5.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

56 extracted references · 6 linked inside Pith

  1. [1]

    arXiv preprint arXiv:2505.15814 (2025) 3

    Arrigoni, F.: A taxonomy of structure from motion methods. arXiv preprint arXiv:2505.15814 (2025) 3

    Pith/arXiv arXiv 2025

  2. [2]

    In: European Conference on Computer Vision

    Avetisyan, A., Xie, C., Howard-Jenkins, H., Yang, T.Y., Aroudj, S., Patra, S., Zhang, F., Frost, D., Holland, L., Orme, C., et al.: Scenescript: Reconstructing scenes with an autoregressive structured language model. In: European Conference on Computer Vision. pp. 247–263. Springer (2024) 10

    2024

  3. [3]

    IEEE transactions on robotics37(6), 1874–1890 (2021) 2, 3

    Campos, C., Elvira, R., Rodríguez, J.J.G., Montiel, J.M., Tardós, J.D.: Orb-slam3: An accurate open-source library for visual, visual–inertial, and multimap slam. IEEE transactions on robotics37(6), 1874–1890 (2021) 2, 3

    2021

  4. [4]

    arXiv preprint arXiv:2406.06521 (2024) 3

    Chen, D., Li, H., Ye, W., Wang, Y., Xie, W., Zhai, S., Wang, N., Liu, H., Bao, H., Zhang, G.: Pgsr: Planar-based gaussian splatting for efficient and high-fidelity surface reconstruction. arXiv preprint arXiv:2406.06521 (2024) 3

    Pith/arXiv arXiv 2024

  5. [5]

    In: European Conference on Computer Vision

    Chen, Y., Xu, H., Zheng, C., Zhuang, B., Pollefeys, M., Geiger, A., Cham, T.J., Cai, J.: Mvsplat: Efficient 3d gaussian splatting from sparse multi-view images. In: European Conference on Computer Vision. pp. 370–386. Springer (2024) 3

    2024

  6. [6]

    Advances in Neural Information Processing Systems37, 107064–107086 (2024) 3

    Chen, Y., Zheng, C., Xu, H., Zhuang, B., Vedaldi, A., Cham, T.J., Cai, J.: Mvs- plat360: Feed-forward 360 scene synthesis from sparse views. Advances in Neural Information Processing Systems37, 107064–107086 (2024) 3

    2024

  7. [7]

    In: Proceedings of the IEEE conference on computer vision and pattern recognition

    Dai,A.,Chang,A.X.,Savva,M.,Halber,M.,Funkhouser,T.,Nießner,M.:Scannet: Richly-annotated 3d reconstructions of indoor scenes. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 5828–5839 (2017) 10

    2017

  8. [8]

    arXiv preprint arXiv:2508.17972 (2025) 11

    Deng,J.,Li,H.,Xie,T.,Ren,W.,Zhang,Q.,Tan,P.,Guo,X.:Sail-recon:Largesfm by augmenting scene regression with localization. arXiv preprint arXiv:2508.17972 (2025) 11

    Pith/arXiv arXiv 2025

  9. [9]

    arXiv preprint arXiv:2512.03422 (2025) 4

    Deng, T., Pan, Y., Yuan, S., Li, D., Wang, C., Li, M., Chen, L., Xie, L., Wang, D., Wang, J., Civera, J., Wang, H., Chen, W.: What is the best 3d scene rep- resentation for robotics? from geometric to foundation models. arXiv preprint arXiv:2512.03422 (2025) 4

    arXiv 2025

  10. [10]

    In: Proceedings of the Computer Vision and Pattern Recognition Conference

    Deng, T., Shen, G., Xun, C., Yuan, S., Jin, T., Shen, H., Wang, Y., Wang, J., Wang, H., Wang, D., et al.: Mne-slam: Multi-agent neural slam for mobile robots. In: Proceedings of the Computer Vision and Pattern Recognition Conference. pp. 1485–1494 (2025) 2, 4, 10

    2025

  11. [11]

    In: European conference on computer vision

    Engel, J., Schöps, T., Cremers, D.: Lsd-slam: Large-scale direct monocular slam. In: European conference on computer vision. pp. 834–849. Springer (2014) 2

    2014

  12. [12]

    In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition

    Gong, Z., Li, X., Tosi, F., Han, J., Mattoccia, S., Cai, J., Poggi, M.: Ov3r: Open- vocabulary semantic 3d reconstruction from rgb videos. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 34206– 34216 (2026) 2, 4

    2026

  13. [13]

    In: European Conference on Computer Vision (2026) 3

    Gong, Z., Li, X., Tosi, F., Zhang, Y., Mattoccia, S., Wu, J., Poggi, M.: Dino- slam: Dino-informed rgb-d slam for neural implicit and explicit representations. In: European Conference on Computer Vision (2026) 3

    2026

  14. [14]

    In: 2025 IEEE International Conference on Robotics and Automation (ICRA)

    Gong, Z., Tosi, F., Zhang, Y., Mattoccia, S., Poggi, M.: Hs-slam: Hybrid rep- resentation with structural supervision for improved dense slam. In: 2025 IEEE International Conference on Robotics and Automation (ICRA). pp. 8464–8470. IEEE (2025) 3

    2025

  15. [15]

    In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition

    Guédon, A., Lepetit, V.: Sugar: Surface-aligned gaussian splatting for efficient 3d mesh reconstruction and high-quality mesh rendering. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 5354– 5363 (2024) 3 MAGiSt3R 17

    2024

  16. [16]

    In: European Conference on Computer Vision

    Ha, S., Yeon, J., Yu, H.: Rgbd gs-icp slam. In: European Conference on Computer Vision. pp. 180–197. Springer (2024) 4

    2024

  17. [17]

    Advances in Neural Information Processing Systems36, 39429– 39442 (2023) 2, 4, 10, 13, 14

    Hu, J., Mao, M., Bao, H., Zhang, G., Cui, Z.: Cp-slam: Collaborative neural point- based slam system. Advances in Neural Information Processing Systems36, 39429– 39442 (2023) 2, 4, 10, 13, 14

    2023

  18. [18]

    In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition

    Huang, H., Li, L., Cheng, H., Yeung, S.K.: Photo-slam: Real-time simultaneous localization and photorealistic mapping for monocular stereo and rgb-d cameras. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 21584–21593 (2024) 4

    2024

  19. [19]

    In: Proceedings of the ieee/cvf conference on computer vision and pattern recognition

    Izquierdo, S., Civera, J.: Optimal transport aggregation for visual place recogni- tion. In: Proceedings of the ieee/cvf conference on computer vision and pattern recognition. pp. 17658–17668 (2024) 6

    2024

  20. [20]

    IEEE Robotics and Automation Letters3(4), 2762–2769 (2018) 4

    Karrer, M., Schmuck, P., Chli, M.: Cvi-slam—collaborative visual-inertial slam. IEEE Robotics and Automation Letters3(4), 2762–2769 (2018) 4

    2018

  21. [21]

    In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition

    Keetha, N., Karhade, J., Jatavallabhula, K.M., Yang, G., Scherer, S., Ramanan, D., Luiten, J.: Splatam: Splat track & map 3d gaussians for dense rgb-d slam. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 21357–21366 (2024) 4

    2024

  22. [22]

    ACM Transactions on Graphics42(4), 1–14 (2023) 2, 3

    Kerbl, B., Kopanas, G., Leimkühler, T., Drettakis, G.: 3d gaussian splatting for real-time radiance field rendering. ACM Transactions on Graphics42(4), 1–14 (2023) 2, 3

    2023

  23. [23]

    IEEE Robotics and Automation Letters9(1), 475–482 (2023) 4, 13, 14

    Lajoie, P.Y., Beltrame, G.: Swarm-slam: Sparse decentralized collaborative si- multaneous localization and mapping framework for multi-robot systems. IEEE Robotics and Automation Letters9(1), 475–482 (2023) 4, 13, 14

    2023

  24. [24]

    In: European Conference on Computer Vision

    Leroy, V., Cabon, Y., Revaud, J.: Grounding image matching in 3d with mast3r. In: European Conference on Computer Vision. pp. 71–91. Springer (2024) 4, 10, 12, 13, 14

    2024

  25. [25]

    arXiv preprint arXiv:2507.23677 (2025) 4

    Li, X., Gong, Z., Tosi, F., Poggi, M., Mattoccia, S., Liu, D., Wu, J.: Stereo 3d gaussian splatting slam for outdoor urban scenes. arXiv preprint arXiv:2507.23677 (2025) 4

    Pith/arXiv arXiv 2025

  26. [26]

    In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition

    Li, Z., Müller, T., Evans, A., Taylor, R.H., Unberath, M., Liu, M.Y., Lin, C.H.: Neuralangelo: High-fidelity neural surface reconstruction. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 8456– 8465 (2023) 3

    2023

  27. [27]

    In: Proceedings of the IEEE/CVF in- ternational conference on computer vision

    Lindenberger, P., Sarlin, P.E., Larsson, V., Pollefeys, M.: Pixel-perfect structure- from-motion with featuremetric refinement. In: Proceedings of the IEEE/CVF in- ternational conference on computer vision. pp. 5987–5997 (2021) 2, 3

    2021

  28. [28]

    In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition

    Liu, S., Yu, Y., Pautrat, R., Pollefeys, M., Larsson, V.: 3d line mapping revisited. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 21445–21455 (2023) 2, 3

    2023

  29. [29]

    In: Proceedings of the Computer Vision and Pattern Recognition Conference

    Liu, Y., Dong, S., Wang, S., Yin, Y., Yang, Y., Fan, Q., Chen, B.: Slam3r: Real- time dense scene reconstruction from monocular rgb videos. In: Proceedings of the Computer Vision and Pattern Recognition Conference. pp. 16651–16662 (2025) 2, 4, 10, 11, 12, 13, 14

    2025

  30. [30]

    arXiv preprint arXiv:2505.12549 (2025) 1, 2, 4, 10, 11, 12, 13, 14

    Maggio, D., Lim, H., Carlone, L.: Vggt-slam: Dense rgb slam optimized on the sl (4) manifold. arXiv preprint arXiv:2505.12549 (2025) 1, 2, 4, 10, 11, 12, 13, 14

    Pith/arXiv arXiv 2025

  31. [31]

    In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recog- nition

    Matsuki, H., Murai, R., Kelly, P.H., Davison, A.J.: Gaussian splatting slam. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recog- nition. pp. 18039–18048 (2024) 4 18 Gong et al

    2024

  32. [32]

    In: Eu- ropean Conference on Computer Vision

    Mildenhall, B., Srinivasan, P.P., Tancik, M., Barron, J.T., Ramamoorthi, R., Ng, R.: Nerf: Representing scenes as neural radiance fields for view synthesis. In: Eu- ropean Conference on Computer Vision. pp. 405–421. Springer (2020) 2, 3

    2020

  33. [33]

    In: Proceedings of the Computer Vision and Pattern Recognition Conference

    Murai, R., Dexheimer, E., Davison, A.J.: Mast3r-slam: Real-time dense slam with 3d reconstruction priors. In: Proceedings of the Computer Vision and Pattern Recognition Conference. pp. 16695–16705 (2025) 4, 10, 12, 13, 14

    2025

  34. [34]

    Peng, Z., Shao, T., Liu, Y., Zhou, J., Yang, Y., Wang, J., Zhou, K.: Rtg-slam: Real- time3dreconstructionatscaleusinggaussiansplatting.In:ACMSIGGRAPH2024 Conference Papers. pp. 1–11 (2024) 4

    2024

  35. [35]

    In: Proceedings of the IEEE/CVF international conference on computer vision

    Ranftl, R., Bochkovskiy, A., Koltun, V.: Vision transformers for dense prediction. In: Proceedings of the IEEE/CVF international conference on computer vision. pp. 12179–12188 (2021) 8

    2021

  36. [36]

    In: Proceedings of the IEEE/CVF International Conference on Computer Vision

    Sandström, E., Li, Y., Van Gool, L., Oswald, M.R.: Point-slam: Dense neural point cloud-based slam. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 18433–18444 (2023) 4

    2023

  37. [37]

    Journal of Field Robotics36(4), 763–781 (2019) 4

    Schmuck, P., Chli, M.: Ccm-slam: Robust and efficient centralized collaborative monocular simultaneous localization and mapping for robotic teams. Journal of Field Robotics36(4), 763–781 (2019) 4

    2019

  38. [38]

    In: Proceedings of the IEEE conference on computer vision and pattern recognition

    Schonberger, J.L., Frahm, J.M.: Structure-from-motion revisited. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 4104–4113 (2016) 2, 3

    2016

  39. [39]

    In: ACM siggraph 2006 papers, pp

    Snavely, N., Seitz, S.M., Szeliski, R.: Photo tourism: exploring photo collections in 3d. In: ACM siggraph 2006 papers, pp. 835–846 (2006) 2, 3

    2006

  40. [40]

    In: Proceedings of the IEEE/CVF international conference on computer vision

    Sucar, E., Liu, S., Ortiz, J., Davison, A.J.: imap: Implicit mapping and position- ing in real-time. In: Proceedings of the IEEE/CVF international conference on computer vision. pp. 6229–6238 (2021) 4

    2021

  41. [41]

    In: 2024 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS)

    Sun, L.C., Bhatt, N.P., Liu, J.C., Fan, Z., Wang, Z., Humphreys, T.E., Topcu, U.: Mm3dgs slam: Multi-modal 3d gaussian splatting for slam using vision, depth, and inertial measurements. In: 2024 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS). pp. 10159–10166. IEEE (2024) 4

    2024

  42. [42]

    Advances in neural information processing systems34, 16558–16569 (2021) 2, 3

    Teed, Z., Deng, J.: Droid-slam: Deep visual slam for monocular, stereo, and rgb- d cameras. Advances in neural information processing systems34, 16558–16569 (2021) 2, 3

    2021

  43. [43]

    arXiv preprint arXiv:2402.132554, 1 (2024) 2, 3

    Tosi, F., Zhang, Y., Gong, Z., Sandström, E., Mattoccia, S., Oswald, M.R., Poggi, M.: How nerfs and 3d gaussian splatting are reshaping slam: a survey. arXiv preprint arXiv:2402.132554, 1 (2024) 2, 3

    Pith/arXiv arXiv 2024

  44. [44]

    arXiv preprint arXiv:2408.15235 (2024) 3

    Wang, F., Zhu, Q., Chang, D., Gao, Q., Han, J., Zhang, T., Hartley, R., Pollefeys, M.: Learning-based multi-view stereo: A survey. arXiv preprint arXiv:2408.15235 (2024) 3

    arXiv 2024

  45. [45]

    In: 2025 Interna- tional Conference on 3D Vision (3DV)

    Wang, H., Agapito, L.: 3d reconstruction with spatial memory. In: 2025 Interna- tional Conference on 3D Vision (3DV). pp. 78–89. IEEE (2025) 4

    2025

  46. [46]

    In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition

    Wang, H., Wang, J., Agapito, L.: Co-slam: Joint coordinate and sparse parametric encodings for neural real-time slam. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 13293–13302 (2023) 4

    2023

  47. [47]

    In: Proceedings of the Computer Vision and Pattern Recognition Conference

    Wang, J., Chen, M., Karaev, N., Vedaldi, A., Rupprecht, C., Novotny, D.: Vggt: Visual geometry grounded transformer. In: Proceedings of the Computer Vision and Pattern Recognition Conference. pp. 5294–5306 (2025) 3, 4, 5

    2025

  48. [48]

    In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition

    Wang, S., Leroy, V., Cabon, Y., Chidlovskii, B., Revaud, J.: Dust3r: Geometric 3d vision made easy. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 20697–20709 (2024) 2, 4, 8, 10, 12, 13, 14 MAGiSt3R 19

    2024

  49. [49]

    In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition

    Yan, C., Qu, D., Xu, D., Zhao, B., Wang, Z., Wang, D., Li, X.: Gs-slam: Dense vi- sual slam with 3d gaussian splatting. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 19595–19604 (2024) 4

    2024

  50. [50]

    In: 2022 IEEE In- ternational Symposium on Mixed and Augmented Reality (ISMAR)

    Yang, X., Li, H., Zhai, H., Ming, Y., Liu, Y., Zhang, G.: Vox-fusion: Dense tracking and mapping with voxel-based neural implicit representation. In: 2022 IEEE In- ternational Symposium on Mixed and Augmented Reality (ISMAR). pp. 499–507. IEEE (2022) 4

    2022

  51. [51]

    In: Proceedings of the IEEE/CVF International Conference on Computer Vision

    Yeshwanth, C., Liu, Y.C., Nießner, M., Dai, A.: Scannet++: A high-fidelity dataset of 3d indoor scenes. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 12–22 (2023) 10

    2023

  52. [52]

    In: Proceedings of the Computer Vision and Pattern Recognition Conference

    Yugay, V., Gevers, T., Oswald, M.R.: Magic-slam: Multi-agent gaussian globally consistent slam. In: Proceedings of the Computer Vision and Pattern Recognition Conference. pp. 6741–6750 (2025) 2, 4, 9, 10, 11, 13, 14

    2025

  53. [53]

    In: Proceedings of the IEEE/CVF International Conference on Computer Vision

    Zhang,Y.,Tosi,F.,Mattoccia,S.,Poggi,M.:Go-slam:Globaloptimizationforcon- sistent 3d instant reconstruction. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 3727–3737 (2023) 4

    2023

  54. [54]

    arXiv preprint arXiv:2511.19031 (2025) 10, 13, 14

    Zhou, Y., Wu, H.: Multi-agent monocular dense slam with 3d reconstruction priors. arXiv preprint arXiv:2511.19031 (2025) 10, 13, 14

    arXiv 2025

  55. [55]

    In: 2025 International Conference on 3D Vision (3DV)

    Zhu, L., Li, Y., Sandström, E., Huang, S., Schindler, K., Armeni, I.: Loopsplat: Loop closure by registering 3d gaussian splats. In: 2025 International Conference on 3D Vision (3DV). pp. 156–167. IEEE (2025) 4

    2025

  56. [56]

    In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition

    Zhu, Z., Peng, S., Larsson, V., Xu, W., Bao, H., Cui, Z., Oswald, M.R., Pollefeys, M.: Nice-slam: Neural implicit scalable encoding for slam. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 12786– 12796 (2022) 4

    2022