Pith. sign in

REVIEW 3 major objections 6 minor 2 cited by

SLAM-Former packs incremental tracking and global map refinement into one transformer, and claims this dual role lets the two halves improve each other, beating state-of-the-art dense SLAM on reconstruction while staying competitive on traj

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 ↗

SLAM-Former unifies online tracking, keyframing, and global map refinement in one transformer and reports the best dense reconstruction accuracy among compared methods on TUM, 7-Scenes, and Replica.

T0 review reviewed 2026-08-04 challenge →

load-bearing objection Credible new SLAM architecture with striking reconstruction gains, but the core training-to-inference equivalence for the backend cache handback is not actually demonstrated. the 3 major comments →

arxiv 2509.16909 v2 pith:TRTOSQXJ submitted 2025-09-21 cs.CV cs.RO

SLAM-Former: Putting SLAM into One Transformer

Yijun Yuan , Zhuoguang Chen , Kenan Li , Weibang Wang , Minghui Qin , Zhijian Fang , Weicheng Zheng , Hang Zhao This is my paper
classification cs.CV cs.RO
keywords SLAMtransformerdense monocular SLAMglobal refinementKV cacheloop closure3D reconstructioncamera tracking
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

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

The reading

SLAM-Former is a single transformer trained to handle both the frontend and backend of a visual SLAM system: it incrementally tracks new frames and builds a map in real time, and it periodically re-processes all stored map tokens with full attention to correct drift and enforce global consistency. The paper claims these two roles are mutually beneficial because the backend's refined key-value cache is handed back to the frontend, so future frames are processed in the context of a corrected global map. The authors train the model in three modes—causal attention for the frontend, full attention for the backend, and a mixed mode that couples them—and evaluate it on standard dense monocular SLAM benchmarks. If the claims hold, it means a single neural network can replace the traditional multi-module SLAM pipeline, eliminating the need for a separate loop-closure module and explicit pose-graph optimization.

Core claim

The paper's central claim is that one transformer can carry a complete visual SLAM system—incremental frontend tracking and mapping plus global backend refinement—and that alternating between the two makes each better. The backend runs full attention over all map tokens, which the authors say is equivalent to performing loop detection on a dense factor graph, and its resulting KV cache is copied back to the frontend so that subsequent frames are processed against the refined global structure. The frontend in turn supplies the backend with sequentially ordered intermediate geometry. Experiments on TUM RGB-D, 7-Scenes, and Replica show reconstruction errors roughly halved compared with strong

What carries the argument

SLAM-Former is built on a shared transformer backbone with map tokens serving as an implicit neural representation of the scene. It operates in three training modes: Mode 1 uses causal attention for the frontend (with full attention on the first two frames to avoid a fixed reference), Mode 3 uses full attention for the backend to refine all map tokens, and Mode 2 uses mixed attention to train frontend and backend together so the frontend can condition on a backend-refined KV cache during inference. The cache handback in Eq. (5)—replacing the frontend's stored KV caches with the backend's refined ones—is the mechanism by which global refinement propagates back into incremental tracking.

Load-bearing premise

The training scheme assumes that training with mixed attention (Mode 2) behaves the same as conditioning the frontend on the backend-refined KV cache during inference, but no experiment directly verifies that this equivalence holds under the alternating execution schedule.

What would settle it

Train SLAM-Former with only Mode 1 (frontend) and Mode 3 (backend), but at inference still perform the backend cache handback; if the resulting tracking and reconstruction accuracy matches the full three-mode model, then Mode 2's mixed attention is not actually load-bearing, and the claimed mutual promotion is an artifact of the architecture rather than the training procedure.

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

If this is right

  • If the claims hold, dense monocular SLAM can be run as a single feed-forward transformer, removing the need for separate loop detection, bundle adjustment, or pose-graph optimization modules.
  • The mutual-promotion design should reduce drift in long sequences because the backend periodically corrects the global structure and the frontend then works from that corrected context.
  • The method is real-time (>10 Hz) on a single consumer GPU while simultaneously improving reconstruction accuracy, which could make it practical for robotics and AR applications.
  • Because the backend uses full attention over all map tokens, the approach naturally handles loop closures without explicit keyframe matching or geometric verification.
  • The training scheme—three modes in one iteration—shows a path for a single model to learn both incremental and global optimization behaviors without architectural changes.

Where Pith is reading between the lines

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

  • If the equivalence between Mode 2 mixed attention and inference-time backend-refined KV conditioning is genuinely tight, the same alternating frontend-backend trick might transfer to other streaming 3D reconstruction models, allowing them to add global consistency without training a second network.
  • The O(n^2) cost of full attention in the backend could be the main scaling bottleneck; the authors mention sparse attention or token merging as future work, so a testable extension would be replacing full attention with a sparse graph and measuring whether the consistency gains survive.
  • The method's reliance on map tokens as implicit geometry suggests that explicit uncertainty or confidence weighting (already present in the loss via Σ*) might be leveraged to decide when to trigger the backend, rather than a fixed keyframe count.
  • Training was initialized from Pi3 pre-trained weights; a natural next question is how much of the improvement comes from the pre-trained geometry prior versus the frontend-backend co-training, which could be tested by ablating the initialization.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

3 major / 6 minor

Summary. The paper introduces SLAM-Former, a single transformer that integrates a SLAM frontend (incremental monocular tracking and mapping) and a global-refinement backend in one architecture. The frontend processes keyframes causally, storing KV caches; the backend periodically refines all map tokens with full attention and hands the refined caches back to the frontend (Eq. (5)). Three training modes are proposed: Mode 1 trains the frontend with causal attention, Mode 2 trains a 'mixed attention' configuration to jointly run backend refinement and frontend inference, and Mode 3 trains pure backend refinement. Experiments on TUM RGB-D, 7-Scenes, and Replica report competitive or state-of-the-art ATE and strong reconstruction accuracy/completeness. The paper claims that the alternating frontend/backend execution leads to mutual promotion and that the backend is equivalent to loop detection on a dense factor graph.

Significance. If the central mechanism holds, SLAM-Former is a meaningful conceptual advance: it replaces a multi-module SLAM pipeline with a single transformer whose frontend and backend share weights and communicate through KV caches. The reported reconstruction gains, e.g., 0.017 m accuracy on 7-Scenes versus >0.05 m for baselines, are substantial. The evaluation is on benchmarks not used in training, and the three-mode training protocol is clearly described at a high level. However, the headline mutual-promotion claim rests on an asserted equivalence between Mode 2 training and the inference-time cache handback; the paper provides no experiment that isolates this equivalence. The backend-to-frontend direction is supported by ablations, but the frontend-to-backend direction is only qualitatively illustrated. These gaps do not disprove the system, but they leave the mechanism underlying the reported gains underspecified.

major comments (3)
  1. [Sec. 3.4, Eq. (5)] The paper states that Mode 2's mixed attention is 'equivalent to conditioning on the backend-refined KV cache' and then uses Eq. (5) to hand the backend cache C_M to the frontend at inference. This equivalence is load-bearing: the frontend is supposed to operate against a frozen, backend-refined representation, not against map tokens whose states were computed jointly with the current frame. The text does not specify the exact attention mask for Mode 2, and Figure 3(b) gives only a schematic. If Mode 2 trains the network with a single forward pass in which map tokens are refined while image tokens attend to those same refined tokens, the model can exploit the jointly computed, 'future-aware' map state. At inference, the cache is frozen after a separate backend run, so Eq. (5) may not reproduce the training distribution. Table 6 does not resolve this, because it compares whole configurati
  2. [Sec. 4.4, Fig. 6] The 'backend benefits from the frontend' half of the mutual-promotion claim is not established by a controlled ablation. The paper addresses this question by comparing SLAM-Former with running VGGT or Pi3 on all keyframe images without sequential information (Fig. 6). This is a comparison against different model families, not an ablation of SLAM-Former itself. It does not show whether SLAM-Former's backend benefits from its own frontend's ordering and initial estimates, as opposed to the backend simply being a more capable module. Table 6 varies only the presence of backend components in the full system; it does not test a row where the backend receives inputs without frontend-provided order. Please add an experiment that isolates the frontend's contribution to the backend, e.g., by running the SLAM-Former backend on unordered or shuffled keyframes with the same architecture and initiali
  3. [Table 4] The Replica reconstruction comparison mixes numbers from different sources and protocols. The text says the evaluation follows the SLAM3R protocol, but Table 4 lists results 'reported in NICER-SLAM' for DROID-SLAM* and NICER-SLAM*, and results 'from SLAM3R' for Spann3R and SLAM3R. If these numbers were produced under a different sampling, alignment, or evaluation protocol, the claim of out-performing all baselines on Replica reconstruction is not directly supported. Please either recompute all baselines under the same protocol or clearly state which protocol each reported number follows. Relatedly, CUT3R+ and StreamVGGT+ are evaluated using the authors' keyframe selection; this should be discussed as a possible advantage, since the keyframe criterion is part of the proposed system.
minor comments (6)
  1. [Sec. 3.5] The backend refresh interval T is introduced but its value is never reported. Please state the value used in the experiments and whether it is tuned per dataset.
  2. [Sec. 3.4] The exact attention pattern of Mode 2 is ambiguous. Please provide a precise mask matrix or pseudocode. The phrase 'causal attention that is equivalent to conditioning on the backend-refined KV cache' is not a formal specification.
  3. [Sec. 1] The statement that the backend 'is equivalent to processing loop detection on a dense factor graph' is a conceptual claim without a formal or experimental justification. Consider softening or providing an analysis (e.g., showing the attention weights correspond to loop-closure constraints).
  4. [Sec. 4.4] The abbreviations MB and EB are used without definition. Please define them (e.g., 'middle backend' and 'end backend') and explain how they differ architecturally from the standard backend described in Sec. 3.3.
  5. [Eq. (1)] The notation F_t = f_fn(I_t){C_k}_{k in S} is unclear: it should be explicit whether f_fn takes the KV cache as a conditioning argument or whether the cache is internal state. Clarify the functional signature.
  6. [Tables 1-3] All reported numbers appear to be single-run results. For a deep SLAM system, it would strengthen the paper to report variance across seeds or at least state that single-run evaluation is standard for this benchmark. This is a presentation issue, not a correctness blocker.

Circularity Check

0 steps flagged

No significant circularity: SLAM-Former is an empirical system evaluated on held-out benchmarks; the frontend/backend mechanism is not a relabeled fit and the self-citations are background.

full rationale

The paper's central claims are system-level empirical results. The backend refinement (Eq. 4) and cache handback (Eq. 5) are architectural mechanisms, not quantities fitted to the evaluation metrics. The evaluations on TUM RGB-D, 7-Scenes, and Replica use datasets that are not in the training list (ARKitScenes, ScanNet, ScanNet++, HyperSim, BlendedMVS, MegaDepth, MVS-Synth), so the reported ATE/Chamfer improvements are not forced by construction. The only author self-citations are SceneFactory [6] and LONG3R [27] in the related-work survey; they are not used to justify any performance, uniqueness, or theoretical claim, so they are not load-bearing. One passage in Sec. 3.4 asserts that Mode 2's mixed attention 'is equivalent to conditioning on the backend-refined KV cache,' and Eq. (5) then hands the backend cache to the frontend. This is a load-bearing architectural assumption whose training/inference match is not isolated by an ablation, but it is an unverified equivalence or missing-support concern, not a circular reduction: the paper does not define Mode 2's mask in terms of the inference cache, and no fitted parameter is renamed as a prediction. The Limitations section (end of Sec. 5) explicitly acknowledges the O(n^2) cost of full attention and the lack of a local frontend mode; these are efficiency/design limitations, not circular steps. Overall, the derivation chain is self-contained against external benchmarks, so the circularity score is 0.

Axiom & Free-Parameter Ledger

5 free parameters · 4 axioms · 0 invented entities

No physical entities are invented. The central claim rests on architectural equivalences and hand-set hyperparameters listed above; the benchmark results are external, so the system is not circular, but its reproducibility depends on unreported parameters.

free parameters (5)
  • camera loss weight lambda = 100
    Hand-set in L = L_depth + L_pmap + lambda * L_cam (Eq. 7); controls pose supervision strength.
  • backend loss weight beta = 10
    Hand-set in L_all = L1 + L2 + beta * L3 (Sec. 3.4).
  • confidence regularization alpha = not reported
    Appears in -alpha * log(Sigma*) in depth and pointmap losses; value not given.
  • keyframe translation threshold tau = not reported
    Keyframe detection in Sec. 3.2 compares relative pose to threshold tau; value not given.
  • backend refresh interval T = not reported
    Sec. 3.5 triggers the backend after every T keyframes; T is not reported.
axioms (4)
  • domain assumption Pi3 pretrained weights provide a transferable geometry prior
    SLAM-Former initializes from Pi3 (Sec. 4.1); if the prior is not transferable to TUM, 7-Scenes, and Replica, the reported results would degrade.
  • domain assumption Backend full attention over all map tokens can correct accumulated drift
    Sec. 3.3 asserts that the global receptive field enables the backend to correct accumulated drift; the ablation shows correlation, not causation.
  • ad hoc to paper Mixed attention training is behaviorally equivalent to conditioning on backend-refined KV caches
    Sec. 3.4 states this equivalence directly; it is the main assumption that makes single-model alternating inference work.
  • domain assumption Ground-truth depth and pointmaps in the training datasets are sufficient supervision
    The loss in Eq. (7) supervises per-frame depth and local pointmaps; the paper does not analyze label noise or dataset bias.

reviewed 2026-08-04 · how reviews work

0 comments
Cite this review

Pith. "Pith review of SLAM-Former: Putting SLAM into One Transformer." pith.science (2026). https://pith.science/paper/TRTOSQXJ

@misc{pith2026250916909,
  author       = {Pith},
  title        = {Pith review of: SLAM-Former: Putting SLAM into One Transformer},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/TRTOSQXJ}},
  note         = {Machine review of arXiv:2509.16909}
}
Share X Bluesky LinkedIn Reddit HN
read the original abstract

We present SLAM-Former, a neural approach that integrates full SLAM capabilities into a single transformer. Similar to traditional SLAM systems, SLAM-Former comprises both a frontend and a back-end that operate in tandem. The frontend processes sequential monocular images in real-time for incremental mapping and tracking, while the backend performs global refinement to ensure a geometrically consistent result. This alternating execution allows the frontend and back-end to mutually promote one another, enhancing overall system performance. Comprehensive experimental results demonstrate that SLAM- Former achieves superior or highly competitive performance compared to state-of-the-art dense SLAM methods.

Figures

Figures reproduced from arXiv: 2509.16909 by Hang Zhao, Kenan Li, Minghui Qin, Weibang Wang, Weicheng Zheng, Yijun Yuan, Zhijian Fang, Zhuoguang Chen.

Figure 1
Figure 1. Figure 1: SLAM-Former is a unified Transformer for SLAM. Traditional SLAM employs a multi-model pipeline for frontend [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: The working pipeline of SLAM-Former. The frontend detects keyframes and performs incremental pose and map [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: Three training modes of SLAM-Former. I and M represent the patch-wise image tokens and map tokens of a frame. and represent the layer intermediate and final output. In each mode, either I or M tokens, or both, are fed into the transformer backbone f, which contains L layers of frame attention and various inter-frame attentions. Finally, pose and pointmap are regressed by the heads h. keyframe indices, S = … view at source ↗
Figure 4
Figure 4. Figure 4: Qualitative reconstruction comparison. Note the significant structural errors, such as misalignments, from baseline [PITH_FULL_IMAGE:figures/full_fig_p007_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: Qualitative reconstruction comparison with and without backend assistance. The first row displays the frontend [PITH_FULL_IMAGE:figures/full_fig_p008_5.png] view at source ↗
Figure 6
Figure 6. Figure 6: Qualitative reconstruction comparison on ICL-NUIM ofkt1. The results from VGGT, Pi3, and ours are displayed [PITH_FULL_IMAGE:figures/full_fig_p009_6.png] view at source ↗

discussion (0)

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

Forward citations

Cited by 2 Pith papers

Reviewed papers in the Pith corpus that reference this work. Sorted by Pith novelty score.

  1. Efficient Feature-Free Initialization for Monocular Visual-Inertial Systems Using a Feed-Forward 3D Model

    cs.RO 2026-05 unverdicted novelty 7.0

    A feature-free monocular VINS initialization method that uses feed-forward 3D model point cloud predictions achieves over 90% success rate with under 1.2 seconds of data and performs robustly in degraded environments.

  2. BA-T: An Iterative Transformer for Two-View Bundle Adjustment

    cs.CV 2026-06 unverdicted novelty 6.0

    BA-T is an iterative Transformer that implements bundle adjustment as a repeatable lightweight layer to progressively refine pose and geometry predictions in two-view 3D reconstruction while using far fewer decoder pa...

Reference graph

Works this paper leans on

44 extracted references · cited by 2 Pith papers

  1. [1]

    Orb- slam: A versatile and accurate monocular slam system,

    R. Mur-Artal, J. M. M. Montiel, and J. D. Tardos, “Orb- slam: A versatile and accurate monocular slam system,” IEEE Transactions on Robotics, vol. 31, p. 1147–1163, Oct. 2015. 9

    2015

  2. [2]

    Lsd-slam: Large-scale direct monocular slam,

    J. Engel, T. Sch ¨ops, and D. Cremers, “Lsd-slam: Large-scale direct monocular slam,” inEuropean conference on com- puter vision, pp. 834–849, Springer, 2014

    2014

  3. [3]

    One billion points in the cloud–an octree for efficient processing of 3d laser scans,

    J. Elseberg, D. Borrmann, and A. N ¨uchter, “One billion points in the cloud–an octree for efficient processing of 3d laser scans,”ISPRS Journal of Photogrammetry and Remote Sensing, vol. 76, pp. 76–88, 2013

    2013

  4. [4]

    Kinectfusion: Real-time dense surface map- ping and tracking,

    R. A. Newcombe, S. Izadi, O. Hilliges, D. Molyneaux, D. Kim, A. J. Davison, P. Kohi, J. Shotton, S. Hodges, and A. Fitzgibbon, “Kinectfusion: Real-time dense surface map- ping and tracking,” in2011 10th IEEE International Sympo- sium on Mixed and Augmented Reality, pp. 127–136, 2011

    2011

  5. [5]

    Droid-slam: Deep visual slam for monocular, stereo, and rgb-d cameras,

    Z. Teed and J. Deng, “Droid-slam: Deep visual slam for monocular, stereo, and rgb-d cameras,” inAdvances in Neural Information Processing Systems(M. Ranzato, A. Beygelzimer, Y . Dauphin, P. Liang, and J. W. Vaughan, eds.), vol. 34, pp. 16558–16569, Curran Associates, Inc., 2021

    2021

  6. [6]

    Scenefactory: A workflow-centric and unified framework for incremental scene modeling,

    Y . Yuan, M. Bleier, and A. N ¨uchter, “Scenefactory: A workflow-centric and unified framework for incremental scene modeling,” 2025

    2025

  7. [7]

    Mast3r-slam: Real-time dense slam with 3d reconstruction priors,

    R. Murai, E. Dexheimer, and A. J. Davison, “Mast3r-slam: Real-time dense slam with 3d reconstruction priors,” 2025

    2025

  8. [8]

    Vggt-slam: Dense rgb slam optimized on the sl(4) manifold,

    D. Maggio, H. Lim, and L. Carlone, “Vggt-slam: Dense rgb slam optimized on the sl(4) manifold,” 2025

    2025

  9. [9]

    Dust3r: Geometric 3d vision made easy,

    S. Wang, V . Leroy, Y . Cabon, B. Chidlovskii, and J. Revaud, “Dust3r: Geometric 3d vision made easy,” 2024

    2024

  10. [10]

    Vggt: Visual geometry grounded transformer,

    J. Wang, M. Chen, N. Karaev, A. Vedaldi, C. Rupprecht, and D. Novotny, “Vggt: Visual geometry grounded transformer,” 2025

    2025

  11. [11]

    Streaming 4d visual geometry transformer,

    D. Zhuo, W. Zheng, J. Guo, Y . Wu, J. Zhou, and J. Lu, “Streaming 4d visual geometry transformer,” 2025

    2025

  12. [12]

    Stream3r: Scalable sequen- tial 3d reconstruction with causal transformer,

    Y . Lan, Y . Luo, F. Hong, S. Zhou, H. Chen, Z. Lyu, S. Yang, B. Dai, C. C. Loy, and X. Pan, “Stream3r: Scalable sequen- tial 3d reconstruction with causal transformer,” 2025

    2025

  13. [13]

    Nicer-slam: Neural implicit scene encoding for rgb slam,

    Z. Zhu, S. Peng, V . Larsson, Z. Cui, M. R. Oswald, A. Geiger, and M. Pollefeys, “Nicer-slam: Neural implicit scene encoding for rgb slam,” 2023

    2023

  14. [14]

    Codeslam - learning a compact, optimisable representation for dense visual slam,

    M. Bloesch, J. Czarnowski, R. Clark, S. Leutenegger, and A. J. Davison, “Codeslam - learning a compact, optimisable representation for dense visual slam,” 2019

    2019

  15. [15]

    Deepfactors: Real-time probabilistic dense monocu- lar slam,

    J. Czarnowski, T. Laidlow, R. Clark, and A. J. Davi- son, “Deepfactors: Real-time probabilistic dense monocu- lar slam,”IEEE Robotics and Automation Letters, vol. 5, p. 721–728, Apr. 2020

    2020

  16. [16]

    Mvsnet: Depth inference for unstructured multi-view stereo,

    Y . Yao, Z. Luo, S. Li, T. Fang, and L. Quan, “Mvsnet: Depth inference for unstructured multi-view stereo,” 2018

    2018

  17. [17]

    Tandem: Tracking and dense mapping in real-time using deep multi- view stereo,

    L. Koestler, N. Yang, N. Zeller, and D. Cremers, “Tandem: Tracking and dense mapping in real-time using deep multi- view stereo,” inConference on Robot Learning, pp. 34–45, PMLR, 2022

    2022

  18. [18]

    Nerf: Representing scenes as neural radiance fields for view synthesis,

    B. Mildenhall, P. P. Srinivasan, M. Tancik, J. T. Barron, R. Ramamoorthi, and R. Ng, “Nerf: Representing scenes as neural radiance fields for view synthesis,” 2020

    2020

  19. [19]

    3d gaussian splatting for real-time radiance field rendering,

    B. Kerbl, G. Kopanas, T. Leimk ¨uhler, and G. Drettakis, “3d gaussian splatting for real-time radiance field rendering,” 2023

    2023

  20. [20]

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

    A. Rosinol, J. J. Leonard, and L. Carlone, “Nerf-slam: Real- time dense monocular slam with neural radiance fields,” 2022

    2022

  21. [21]

    Gs-slam: Dense visual slam with 3d gaussian splatting,

    C. Yan, D. Qu, D. Xu, B. Zhao, Z. Wang, D. Wang, and X. Li, “Gs-slam: Dense visual slam with 3d gaussian splatting,” 2024

    2024

  22. [22]

    Grounding image matching in 3d with mast3r,

    V . Leroy, Y . Cabon, and J. Revaud, “Grounding image matching in 3d with mast3r,” 2024

    2024

  23. [23]

    Fast3r: Towards 3d recon- struction of 1000+ images in one forward pass,

    J. Yang, A. Sax, K. J. Liang, M. Henaff, H. Tang, A. Cao, J. Chai, F. Meier, and M. Feiszli, “Fast3r: Towards 3d recon- struction of 1000+ images in one forward pass,” 2025

    2025

  24. [24]

    π 3: Scalable permutation- equivariant visual geometry learning,

    Y . Wang, J. Zhou, H. Zhu, W. Chang, Y . Zhou, Z. Li, J. Chen, J. Pang, C. Shen, and T. He, “π 3: Scalable permutation- equivariant visual geometry learning,” 2025

    2025

  25. [25]

    3d reconstruction with spatial memory,

    H. Wang and L. Agapito, “3d reconstruction with spatial memory,” 2024

    2024

  26. [26]

    Continuous 3d perception model with per- sistent state,

    Q. Wang, Y . Zhang, A. Holynski, A. A. Efros, and A. Kanazawa, “Continuous 3d perception model with per- sistent state,” 2025

    2025

  27. [27]

    Long3r: Long sequence streaming 3d reconstruction,

    Z. Chen, M. Qin, T. Yuan, Z. Liu, and H. Zhao, “Long3r: Long sequence streaming 3d reconstruction,” 2025

    2025

  28. [28]

    Deepv2d: Video to depth with differ- entiable structure from motion,

    Z. Teed and J. Deng, “Deepv2d: Video to depth with differ- entiable structure from motion,” 2020

    2020

  29. [29]

    Deep patch visual slam,

    L. Lipson, Z. Teed, and J. Deng, “Deep patch visual slam,” 2024

    2024

  30. [30]

    Go-slam: Global optimization for consistent 3d instant reconstruction,

    Y . Zhang, F. Tosi, S. Mattoccia, and M. Poggi, “Go-slam: Global optimization for consistent 3d instant reconstruction,” 2023

    2023

  31. [31]

    A benchmark for the evaluation of rgb-d slam sys- tems,

    J. Sturm, N. Engelhard, F. Endres, W. Burgard, and D. Cre- mers, “A benchmark for the evaluation of rgb-d slam sys- tems,” inProc. of the International Conference on Intelligent Robot Systems (IROS), Oct. 2012

    2012

  32. [32]

    Real- time rgb-d camera relocalization,

    B. Glocker, S. Izadi, J. Shotton, and A. Criminisi, “Real- time rgb-d camera relocalization,” inInternational Sympo- sium on Mixed and Augmented Reality (ISMAR), IEEE, Oc- tober 2013

    2013

  33. [33]

    Slam3r: Real-time dense scene reconstruction from monocular rgb videos,

    Y . Liu, S. Dong, S. Wang, Y . Yin, Y . Yang, Q. Fan, and B. Chen, “Slam3r: Real-time dense scene reconstruction from monocular rgb videos,” 2025

    2025

  34. [34]

    The replica dataset: A digital replica of indoor spaces,

    J. Straub, T. Whelan, L. Ma, Y . Chen, E. Wijmans, S. Green, J. J. Engel, R. Mur-Artal, C. Ren, S. Verma, A. Clark- son, M. Yan, B. Budge, Y . Yan, X. Pan, J. Yon, Y . Zou, K. Leon, N. Carter, J. Briales, T. Gillingham, E. Mueggler, L. Pesqueira, M. Savva, D. Batra, H. M. Strasdat, R. D. Nardi, M. Goesele, S. Lovegrove, and R. Newcombe, “The replica datas...

    2019

  35. [35]

    Arkitscenes: A diverse real-world dataset for 3d in- door scene understanding using mobile rgb-d data,

    G. Baruch, Z. Chen, A. Dehghan, T. Dimry, Y . Feigin, P. Fu, T. Gebauer, B. Joffe, D. Kurz, A. Schwartz, and E. Shul- man, “Arkitscenes: A diverse real-world dataset for 3d in- door scene understanding using mobile rgb-d data,” 2022

    2022

  36. [36]

    Scannet: Richly-annotated 3d reconstruc- tions of indoor scenes,

    A. Dai, A. X. Chang, M. Savva, M. Halber, T. Funkhouser, and M. Nießner, “Scannet: Richly-annotated 3d reconstruc- tions of indoor scenes,” inProc. Computer Vision and Pat- tern Recognition (CVPR), IEEE, 2017

    2017

  37. [37]

    Scan- net++: A high-fidelity dataset of 3d indoor scenes,

    C. Yeshwanth, Y .-C. Liu, M. Nießner, and A. Dai, “Scan- net++: A high-fidelity dataset of 3d indoor scenes,” inPro- ceedings of the International Conference on Computer Vi- sion (ICCV), 2023. 10

    2023

  38. [38]

    Hy- persim: A photorealistic synthetic dataset for holistic indoor scene understanding,

    M. Roberts, J. Ramapuram, A. Ranjan, A. Kumar, M. A. Bautista, N. Paczan, R. Webb, and J. M. Susskind, “Hy- persim: A photorealistic synthetic dataset for holistic indoor scene understanding,” inInternational Conference on Com- puter Vision (ICCV) 2021, 2021

    2021

  39. [39]

    Blendedmvs: A large-scale dataset for generalized multi-view stereo networks,

    Y . Yao, Z. Luo, S. Li, J. Zhang, Y . Ren, L. Zhou, T. Fang, and L. Quan, “Blendedmvs: A large-scale dataset for generalized multi-view stereo networks,” 2020

    2020

  40. [40]

    Megadepth: Learning single-view depth prediction from internet photos,

    Z. Li and N. Snavely, “Megadepth: Learning single-view depth prediction from internet photos,” inComputer Vision and Pattern Recognition (CVPR), 2018

    2018

  41. [41]

    Deepmvs: Learning multi-view stereopsis,

    P.-H. Huang, K. Matzen, J. Kopf, N. Ahuja, and J.-B. Huang, “Deepmvs: Learning multi-view stereopsis,” in IEEE Conference on Computer Vision and Pattern Recog- nition (CVPR), 2018

    2018

  42. [42]

    Orb-slam3: An accurate open-source li- brary for visual, visual–inertial, and multimap slam,

    C. Campos, R. Elvira, J. J. G. Rodriguez, J. M. M. Montiel, and J. D. Tardos, “Orb-slam3: An accurate open-source li- brary for visual, visual–inertial, and multimap slam,”IEEE Transactions on Robotics, vol. 37, p. 1874–1890, Dec. 2021

    2021

  43. [43]

    Dense rgb slam with neural implicit maps,

    H. Li, X. Gu, W. Yuan, L. Yang, Z. Dong, and P. Tan, “Dense rgb slam with neural implicit maps,” 2023

    2023

  44. [44]

    A benchmark for rgb-d visual odometry, 3d reconstruction and slam,

    A. Handa, T. Whelan, J. McDonald, and A. J. Davison, “A benchmark for rgb-d visual odometry, 3d reconstruction and slam,” in2014 IEEE international conference on Robotics and automation (ICRA), pp. 1524–1531, IEEE, 2014. 11

    2014

This paper was first reviewed by deepseek-v4-flash on August 4, 2026.