Pith. sign in

REVIEW 2 major objections 5 minor 52 references

Under distribution shift, learning to propose stereo correspondences and their uncertainty, then verifying them geometrically, outperforms learning more of the VIO pipeline.

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 ↗

Restricting learning to stereo correspondence and uncertainty, with geometric verification and covariance weighting in an explicit VIO backend, improves OOD trajectory accuracy and success rate over classical and learning-dominant baselines.

T0 review reviewed 2026-08-01 challenge →

load-bearing objection Solid hybrid VIO paper, but the headline claim overreaches because every learning-dominant baseline is monocular/no-IMU while the proposed system is stereo-inertial. the 2 major comments →

arxiv 2607.17956 v1 pith:6JLVOADV submitted 2026-07-20 cs.RO

Does Robust VIO Need More Learning? Geometry-Verified Visual Measurements under Distribution Shift

Yangyang Ning , Shu Liang , Quanbo Ge , Tianchen Deng , Yuhua Qi , Shenghai Yuan This is my paper
classification cs.RO
keywords visual-inertial odometrydistribution shiftstereo VIOlearned optical flowuncertainty-aware weightinggeometric verificationsliding-window optimization3D Gaussian mapping
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

The paper is trying to establish that robust visual-inertial odometry under out-of-distribution conditions does not require replacing the geometric estimator with learned motion, depth, or structure. It argues that confining learning to dense stereo correspondence generation and uncertainty prediction, while keeping explicit temporal tracking, epipolar verification, and sliding-window inertial optimization, gives the best balance of accuracy and success rate on shifted benchmarks. On standard indoor, dynamic-simulation, and real-world outdoor sequences, the proposed hybrid system achieves best or near-best trajectory error while remaining competitive on an in-domain benchmark. The authors specifically show that learned flow alone is insufficient: performance gains appear only when learned correspondences are filtered by predicted uncertainty, verified with epipolar geometry, and weighted by a converted covariance in the reprojection factors. If true, this would mean that the bottleneck for out-of-distribution robustness is not how much of the pipeline learns, but whether learned measurements are turned into estimator-compatible probabilistic constraints.

Core claim

On the paper's own terms, the central discovery is that the value of learning in VIO lies in proposing visual measurements, not in solving the estimation problem. Using a learned dense optical-flow network to generate stereo correspondences and per-pixel uncertainty, the method samples that dense flow at sparse feature locations, rejects correspondences with high predicted uncertainty or epipolar inconsistency, and inserts survivors into a stereo-inertial sliding-window optimizer as covariance-weighted reprojection factors. The same uncertainty is propagated through triangulation into 3D Gaussian map primitives. Across indoor, dynamic-simulation, and outdoor real-world benchmarks with strong

What carries the argument

The carrying mechanism is a hybrid front-end: a conventional Kanade-Lucas-Tomasi tracker handles temporal association, while a learned dense stereo flow network (SEA-RAFT, defined in one phrase as an efficient recurrent optical-flow model) proposes left-to-right correspondences by bilinearly sampling dense flow at sparse FAST feature locations and emits mixture-of-Laplacian uncertainty. The output is normalized to [0,1], used to reject uncertain matches, and then further gated by a stereo epipolar check. For accepted observations, predicted uncertainty is converted into an image-domain covariance proxy Sigma_pro = Sigma_obs + lambda * sigma_hat^2 * I, which weights the reprojection factor in

Load-bearing premise

The paper's advantage rests on the assumption that the learned flow network's predicted uncertainty, after min-max normalization and scaling by a fixed uncalibrated constant, remains a trustworthy relative measure of observation error under distribution shift; if the network is confidently wrong in a new domain, the geometric and weighting safeguards may not save the estimate.

What would settle it

Run the same pipeline on a shifted sequence where the flow network is deliberately deceived—for example, a low-texture corridor with repetitive vertical edges or heavy fog—and measure whether the uncertainty gate rejects the bad matches. If predicted uncertainty stays low for systematically wrong correspondences that pass the epipolar check, and the trajectory error rises sharply, then the covariance-weighting mechanism ceases to protect the estimator in exactly the case the paper names as a remaining challenge.

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

If this is right

  • If the central claim holds, system designers can gain out-of-distribution robustness by swapping only the stereo matching stage for a learned proposer plus uncertainty, leaving the rest of a classical VIO pipeline unchanged.
  • The ablations imply that simply replacing a hand-crafted matcher with a learned one will not reproduce the result; uncertainty filtering, epipolar verification, and covariance weighting each contribute to the trajectory-error reduction.
  • The propagated uncertainty gives downstream 3D Gaussian mapping a covariance-informed initialization, which the paper shows qualitatively as sharper reconstruction under out-of-distribution conditions.
  • The measured timing—about 44 ms for the learned stereo stage at one iteration—indicates the design can run near 20 Hz on a mid-range GPU, making the robustness gain plausibly deployable.

Where Pith is reading between the lines

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

  • An untested extension is that the same recipe—learned correspondence proposals, geometric verification, and covariance weighting—should transfer to monocular or RGB-D VIO, with the epipolar verifier replaced by multi-view consistency, since the proposed architecture is not intrinsically stereo-specific.
  • Because the covariance proxy is explicitly uncalibrated, one could test whether calibrating the scale factor per domain with a short supervised segment changes the trajectory error; if calibration helps substantially, the gain may be partly in the weighting, and if it does not, the gain may be almost entirely in the uncertainty gating.
  • The paper names a remaining failure case where incorrect correspondences are geometrically plausible; a natural stress test is a low-texture or repetitive-texture scene where wrong matches lie close to epipolar lines, so the geometric gate passes them and the uncertainty gate must carry the full burden.
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

2 major / 5 minor

Summary. The paper proposes a stereo VIO framework that deliberately restricts learning to dense stereo correspondence and uncertainty prediction, using SEA-RAFT only as a correspondence/uncertainty proposer. Temporal tracking remains KLT-based; learned stereo matches are filtered by predicted uncertainty and stereo epipolar consistency; the surviving correspondences enter a sliding-window stereo-inertial estimator through covariance-weighted reprojection factors. The same uncertainty is propagated through stereo triangulation to initialize and weight 3D Gaussian mapping. Experiments on EuRoC, VIODE, and 4Seasons report ATE RMSE and success rates across 3,000+ runs and show that the proposed method achieves the best or near-best accuracy and stability on the shifted VIODE and 4Seasons benchmarks while remaining competitive on EuRoC. Ablations indicate that learned flow alone is not sufficient: the gains come from combining learned correspondence proposals with geometric verification and uncertainty-aware weighting.

Significance. If the central claim holds, the paper is a useful and falsifiable counterpoint to the trend toward learning-dominant VO/VIO: a “minimal-learning” stereo-inertial system with explicit geometric verification can outperform deeper learned estimators under distribution shift. I credit the authors for a broad evaluation across three benchmarks, explicit ablation of the verification and weighting stages, and the honest acknowledgment that the network-derived covariance proxy is not statistically calibrated. The comparison does not currently isolate the amount of learning as the causal variable, however, because the learning-dominant baselines are all monocular without IMU. With that confound removed or the conclusion narrowed, the result would be a meaningful contribution to the robust-VIO literature.

major comments (2)
  1. [§IV-B, Table II] The headline conclusion (“robust generalization does not require a heavy learning pipeline”) is not isolated by the experimental design. The learning-dominant baselines (DROID-VO, DPVO, TartanVO, VGGT-SLAM) are all monocular and do not receive IMU measurements, whereas the proposed system is stereo with IMU and a classical geometric backend. The large improvements on VIODE and 4Seasons could therefore be caused by the sensor configuration and explicit backend rather than by the decision to restrict learning to measurement generation. Please add an appropriate stereo-inertial (or at least stereo, non-IMU) learning-dominant baseline—for example DROID-SLAM in stereo odometry mode, or a stereo learned-depth/flow VO fused with IMU—and, if feasible, a monocular variant of the proposed method. In the meantime, the conclusion should be narrowed to what the comparison actually shows.
  2. [§III-B, §III-C, Eqs. (6)–(9), Table III] The uncertainty mechanism is the key ingredient in the claimed gains, but the parameters that control it are not reported: λ in Eq. (8), ε_net in Eq. (6), and ε_geo in Eq. (7) appear in no table or text, and no sensitivity study is provided. Because Eq. (9) weights the residual by Σ_pro^{-1}, the numerical value of λ can dominate the behavior of the learned observations; without it, the reported ablation gains are not reproducible. The per-image min-max normalization in Eq. (6) also makes σ̂ image-relative, so a fixed ε_net may mean different things on EuRoC, VIODE, and 4Seasons. Please report the values, justify the normalization choice, and show that the conclusion (e.g., the “Full proposed” row of Table III) is stable over a plausible range of λ and ε.
minor comments (5)
  1. [§IV, Tables I–II] All values are five-run averages, but no standard deviations or variance bars are reported. This makes it difficult to judge whether small differences on EuRoC (e.g., 0.03 vs. 0.04 m) are meaningful. Please report variability or state that the runs are deterministic.
  2. [Abstract / §IV Implementation Details] Code and configs are promised only upon acceptance. For reproducibility, the exact SEA-RAFT checkpoint, flow sampling settings, and backend parameter file should be released or described in detail.
  3. [§III-B, Eq. (7)] The epipolar residual uses the essential matrix with “unit direction vectors.” The algebraic residual in Eq. (7) is scale-dependent in its formulation; the standard Sampson error would be clearer and scale-invariant, and would make the threshold ε_geo more interpretable.
  4. [§IV-D, Table III] The discussion mentions “temporal consistency checking” as part of the incremental ablation, but Table III lists only “Epipolar verification.” Clarify whether temporal consistency was used and report its quantitative effect separately if it is part of the claimed contribution.
  5. [§III-D, Fig. 5] The 3D Gaussian mapping result is only qualitative. If uncertainty propagation to mapping is a claimed contribution, a quantitative metric (e.g., rendering PSNR/SSIM or depth accuracy) would substantially strengthen the evaluation.

Circularity Check

0 steps flagged

No significant circularity: the uncertainty-weighted stereo VIO pipeline is an explicit modeling design, and its reported gains are empirical outcomes rather than inputs renamed as predictions.

full rationale

The paper's derivation chain is not circular. The learned uncertainty enters at Eq. (5) as SEA-RAFT's mixture-Laplacian output, is normalized in Eq. (6), and is assembled into a covariance proxy in Eq. (8) as Sigma_net = lambda*sigma_hat^2*I. This covariance is then used in the reprojection factor Eq. (9) and in the triangulation Hessian Eq. (11). This is a modeling choice about how to weight learned observations; the ATE RMSE results in Tables I-III are measured outcomes, not quantities that are equal to the uncertainty by construction. The ablation in Table III explicitly varies the verification and covariance components and shows monotone ATE improvement, which is exactly the kind of independent empirical content that rules out a purely definitional relation. There is no fitted parameter that is later reported as a prediction: lambda is a fixed, unreported scale, and the paper openly states the normalized network uncertainty is 'not treated as a statistically calibrated image-domain covariance.' The references to SEA-RAFT and standard VIO baselines are external and do not form a load-bearing self-citation chain; no uniqueness theorem or prior result by the same authors is invoked to force the design. A skeptical concern that the comparison confounds sensor modality and IMU usage with the amount of learning is a threat to the strength of the empirical conclusion, not a circularity of the derivation. Under the specified standard, no circular step can be quoted, so the appropriate score is 0.

Axiom & Free-Parameter Ledger

3 free parameters · 4 axioms · 0 invented entities

The central claim adds no new physical entities. It rests on four assumptions and three unreported hand-set constants; the lambda scale is the most consequential because it directly shapes the uncertainty-weighted factors that the ablations credit for the gain.

free parameters (3)
  • lambda (learned-flow covariance scale) = not reported
    Eq. (8) sets Sigma_net = lambda * sigma_hat^2 * I; controls how strongly learned uncertainty weights reprojection factors. Value is not given, so the reader cannot tell whether it was tuned per benchmark.
  • epsilon_net (learned-uncertainty threshold) = not reported
    Section III-B: tracks with sigma_hat(z_l^i) < epsilon_net are kept; threshold is hand-chosen and unreported.
  • epsilon_geo (epipolar consistency threshold) = not reported
    Eq. (7): correspondences with |z_l_hat^T E z_r_hat| <= epsilon_geo are kept; threshold is unreported.
axioms (4)
  • domain assumption SEA-RAFT pretrained on TartanAir produces meaningful stereo flow and uncertainty on unseen EuRoC, VIODE, and 4Seasons domains.
    All OOD gains depend on the network transferring; not verified beyond the reported benchmarks.
  • ad hoc to paper Network-predicted uncertainty, after min-max normalization, is a valid relative reliability measure that can be used as a covariance proxy.
    Eqs. (5)-(8); explicitly uncalibrated, yet used both for filtering and weighting.
  • domain assumption Calibrated stereo extrinsics are known and the epipolar constraint is a reliable geometric check.
    Eq. (7) assumes accurate E_Il,Ir.
  • ad hoc to paper Observation noise and learned-flow uncertainty are independent, so covariances add (Sigma_pro = Sigma_obs + Sigma_net).
    Eq. (8); independence is asserted, not justified.

reviewed 2026-08-01 · how reviews work

0 comments
Cite this review

Pith. "Pith review of Does Robust VIO Need More Learning? Geometry-Verified Visual Measurements under Distribution Shift." pith.science (2026). https://pith.science/paper/6JLVOADV

@misc{pith2026260717956,
  author       = {Pith},
  title        = {Pith review of: Does Robust VIO Need More Learning? Geometry-Verified Visual Measurements under Distribution Shift},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/6JLVOADV}},
  note         = {Machine review of arXiv:2607.17956}
}
Share X Bluesky LinkedIn Reddit HN
read the original abstract

Learning is increasingly introduced into visual-inertial odometry (VIO), ranging from learned feature front-ends to learning-dominant motion and geometry estimation. However, learning more of the pipeline does not necessarily improve robustness when deployment conditions differ from the training distribution. This work asks whether robust VIO under distribution shift truly requires deeper learned estimation, or whether learning can be confined to visual measurement generation. We propose a minimal-learning stereo VIO framework in which SEA-RAFT is used only to propose dense stereo correspondences and predict their uncertainty, while temporal tracking, geometric verification, and state estimation remain explicit. Dense flow is sampled at sparse feature locations, filtered using predicted uncertainty and stereo epipolar consistency, and incorporated into a sliding-window stereo-inertial estimator through uncertainty-weighted reprojection factors. The same uncertainty is further propagated through stereo triangulation for downstream anisotropic 3D Gaussian mapping. Experiments on EuRoC, VIODE, and 4Seasons demonstrate accurate and stable estimation under motion blur, dynamic scenes, illumination changes, and large indoor-to-outdoor distribution shifts. Ablations show that learned flow alone is insufficient: the gains arise from combining learned correspondence proposals with geometric verification and uncertainty-aware weighting. These results suggest that, for OOD-robust VIO, carefully integrated learned visual measurements can be more effective than learning a larger fraction of the estimation pipeline. Code and configs for the benchmark will be open-source upon acceptance. A supplementary video is available at https://drive.google.com/file/d/1EVRhOkhanmNXHbQS1Vr80FoEIAYOYOV2/view

Figures

Figures reproduced from arXiv: 2607.17956 by Quanbo Ge, Shenghai Yuan, Shu Liang, Tianchen Deng, Yangyang Ning, Yuhua Qi.

Figure 1
Figure 1. Figure 1: Qualitative positioning of representative geometric, learning-dominant, [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Overview of the proposed hybrid system pipeline. Geometric-verified learned optical flow is utilized for stereo tracking, while the learned uncertainty [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: Qualitative and quantitative comparison of temporal tracking and [PITH_FULL_IMAGE:figures/full_fig_p004_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: Learning-informed stereo triangulation and uncertainty propagation [PITH_FULL_IMAGE:figures/full_fig_p005_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: Qualitative 3D Gaussian mapping results on an out-of-distribution [PITH_FULL_IMAGE:figures/full_fig_p006_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

52 extracted references · 1 linked inside Pith

  1. [1]

    Keyframe-based visual–inertial odometry using nonlinear optimization,

    S. Leuteneggeret al., “Keyframe-based visual–inertial odometry using nonlinear optimization,”Int. J. Robot. Res., 2015

    2015

  2. [2]

    VINS-Mono: A robust and versatile monocular visual– inertial state estimator,

    T. Qinet al., “VINS-Mono: A robust and versatile monocular visual– inertial state estimator,”TRO, 2018

    2018

  3. [3]

    OpenVINS: A research platform for visual–inertial estimation,

    P. Genevaet al., “OpenVINS: A research platform for visual–inertial estimation,” inICRA, 2020

    2020

  4. [4]

    Good features to track,

    J. Shiet al., “Good features to track,” inCVPR, 1994

    1994

  5. [5]

    ORB: An efficient alternative to SIFT or SURF,

    E. Rubleeet al., “ORB: An efficient alternative to SIFT or SURF,” in Proc. IEEE Int. Conf. Comput. Vis. (ICCV), 2011

    2011

  6. [6]

    Bundle adjustment: A modern synthesis,

    B. Triggset al., “Bundle adjustment: A modern synthesis,” inProc. Int. Workshop Vis. Algorithms (IWVA), 2000

    2000

  7. [7]

    T. D. Barfoot,State Estimation for Robotics. Cambridge Univ. Press, 2017

    2017

  8. [8]

    Thrunet al.,Probabilistic Robotics

    S. Thrunet al.,Probabilistic Robotics. MIT Press, 2005

    2005

  9. [9]

    Visual odometry: Part i: The first 30 years and fundamentals,

    F. Fraundorferet al., “Visual odometry: Part i: The first 30 years and fundamentals,”IEEE Robot. Autom. Mag., 2011

    2011

  10. [10]

    Visual odometry: Part ii: Matching, robustness, optimization, and applications,

    F. Fraundorferet al., “Visual odometry: Part ii: Matching, robustness, optimization, and applications,”IEEE Robot. Autom. Mag., 2012

    2012

  11. [11]

    Past, present, and future of simultaneous localization and mapping: Toward the robust-perception age,

    C. Cadenaet al., “Past, present, and future of simultaneous localization and mapping: Toward the robust-perception age,”TRO, 2016

    2016

  12. [12]

    AirSLAM: An efficient and illumination-robust point-line visual SLAM system,

    K. Xuet al., “AirSLAM: An efficient and illumination-robust point-line visual SLAM system,”TRO, 2025

    2025

  13. [13]

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

    Z. Teedet al., “DROID-SLAM: Deep visual SLAM for monocular, stereo, and RGB-D cameras,” inNeurIPS, 2021

    2021

  14. [14]

    Deep patch visual odometry,

    Z. Teedet al., “Deep patch visual odometry,” inNeurIPS, 2023

    2023

  15. [15]

    MASt3R-SLAM: Real-time dense SLAM with 3D reconstruction priors,

    R. Muraiet al., “MASt3R-SLAM: Real-time dense SLAM with 3D reconstruction priors,” inCVPR, 2025

    2025

  16. [16]

    VGGT: Visual geometry grounded transformer,

    J. Wanget al., “VGGT: Visual geometry grounded transformer,” in CVPR, 2025

    2025

  17. [17]

    TartanVO: A generalizable learning-based VO,

    W. Wanget al., “TartanVO: A generalizable learning-based VO,” in CoRL, 2021

    2021

  18. [18]

    Mac-vo: Metrics-aware covariance for learning-based stereo visual odometry,

    Y . Qiuet al., “Mac-vo: Metrics-aware covariance for learning-based stereo visual odometry,” inICRA, 2025

    2025

  19. [19]

    Visual–inertial mapping with non-linear factor recov- ery,

    V . Usenkoet al., “Visual–inertial mapping with non-linear factor recov- ery,”RAL, 2020

    2020

  20. [20]

    SchurVINS: Schur complement-based lightweight visual inertial navigation system,

    Y . Fanet al., “SchurVINS: Schur complement-based lightweight visual inertial navigation system,” inCVPR, 2024

    2024

  21. [21]

    RD-VIO: Robust visual–inertial odometry for mobile augmented reality in dynamic environments,

    J. Liet al., “RD-VIO: Robust visual–inertial odometry for mobile augmented reality in dynamic environments,”IEEE Trans. Vis. Comput. Graph., 2024

    2024

  22. [22]

    Okvis2: Realtime scalable visual-inertial slam with loop closure,

    S. Leutenegger, “Okvis2: Realtime scalable visual-inertial slam with loop closure,”arXiv preprint arXiv:2202.09199, 2022

    Pith/arXiv arXiv 2022

  23. [23]

    DM-VIO: Delayed marginalization visual– inertial odometry,

    L. von Stumberget al., “DM-VIO: Delayed marginalization visual– inertial odometry,”RAL, 2022

    2022

  24. [24]

    RAFT: Recurrent all-pairs field transforms for optical flow,

    Z. Teedet al., “RAFT: Recurrent all-pairs field transforms for optical flow,” inECCV, 2020

    2020

  25. [25]

    SEA-RAFT: Simple, efficient, accurate RAFT for optical flow,

    Y . Wanget al., “SEA-RAFT: Simple, efficient, accurate RAFT for optical flow,” inECCV, 2024

    2024

  26. [26]

    Visual–inertial odometry using optical flow from deep learning,

    J. M. Kanget al., “Visual–inertial odometry using optical flow from deep learning,” inProc. Int. Conf. Inf. Fusion (FUSION), 2024

    2024

  27. [27]

    DeepOF-VIO: A filter-based visual–inertial odometry using deep optical flow,

    J. Kang, “DeepOF-VIO: A filter-based visual–inertial odometry using deep optical flow,”IEEE Sens. J., 2026

    2026

  28. [28]

    D3VO: Deep depth, deep pose and deep uncertainty for monocular visual odometry,

    N. Yanget al., “D3VO: Deep depth, deep pose and deep uncertainty for monocular visual odometry,” inCVPR, 2020

    2020

  29. [29]

    Associating uncertainty with three-dimensional poses for use in estimation problems,

    T. D. Barfootet al., “Associating uncertainty with three-dimensional poses for use in estimation problems,”TRO, 2014

    2014

  30. [30]

    Characterizing the uncertainty of jointly distributed poses in the lie algebra,

    J. G. Mangelsonet al., “Characterizing the uncertainty of jointly distributed poses in the lie algebra,”TRO, 2020

    2020

  31. [31]

    At all costs: A comparison of robust cost functions for camera correspondence outliers,

    K. MacTavishet al., “At all costs: A comparison of robust cost functions for camera correspondence outliers,” inCRV, 2015

    2015

  32. [32]

    Graduated non-convexity for robust spatial perception: From non-minimal solvers to global outlier rejection,

    H. Yanget al., “Graduated non-convexity for robust spatial perception: From non-minimal solvers to global outlier rejection,”RAL, 2020

    2020

  33. [33]

    Consistency analysis and improvement of vision- aided inertial navigation,

    J. A. Heschet al., “Consistency analysis and improvement of vision- aided inertial navigation,”TRO, 2014

    2014

  34. [34]

    On-manifold preintegration for real-time visual– inertial odometry,

    C. Forsteret al., “On-manifold preintegration for real-time visual– inertial odometry,”TRO, 2017

    2017

  35. [35]

    Good feature selection for least squares pose optimiza- tion in VO/VSLAM,

    Y . Zhaoet al., “Good feature selection for least squares pose optimiza- tion in VO/VSLAM,” inIROS, 2018

    2018

  36. [36]

    Good feature matching: Toward accurate, robust VO/VSLAM with low latency,

    Y . Zhaoet al., “Good feature matching: Toward accurate, robust VO/VSLAM with low latency,”TRO, 2020

    2020

  37. [37]

    Information sparsification in visual–inertial odometry,

    J. Hsiunget al., “Information sparsification in visual–inertial odometry,” inIROS, 2018

    2018

  38. [38]

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

    B. Mildenhallet al., “NeRF: Representing scenes as neural radiance fields for view synthesis,” inECCV, 2020

    2020

  39. [39]

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

    L. Koestleret al., “TANDEM: Tracking and dense mapping in real-time using deep multi-view stereo,” inCoRL, 2022

    2022

  40. [40]

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

    B. Kerblet al., “3D gaussian splatting for real-time radiance field rendering,”ACM Trans. Graph., 2023

    2023

  41. [41]

    Maplab: An open framework for research in visual– inertial mapping and localization,

    T. Schneideret al., “Maplab: An open framework for research in visual– inertial mapping and localization,”RAL, 2018

    2018

  42. [42]

    Visual–inertial odometry and active dense reconstruction for mobile robots,

    C. Forster, “Visual–inertial odometry and active dense reconstruction for mobile robots,” Ph.D. dissertation, University of Zurich, 2016

    2016

  43. [43]

    Motion tracking with fixed-lag smoothing: Algo- rithm and consistency analysis,

    T.-C. Dong-Siet al., “Motion tracking with fixed-lag smoothing: Algo- rithm and consistency analysis,” inICRA, 2011

    2011

  44. [44]

    Anysplat: Feed-forward 3d gaussian splatting from unconstrained views,

    L. Jianget al., “Anysplat: Feed-forward 3d gaussian splatting from unconstrained views,”ACM Transactions on Graphics, 2025

    2025

  45. [45]

    The EuRoC micro aerial vehicle datasets,

    M. Burriet al., “The EuRoC micro aerial vehicle datasets,”Int. J. Robot. Res., 2016

    2016

  46. [46]

    VIODE: A simulated dataset to address the challenges of visual–inertial odometry in dynamic environments,

    K. Minodaet al., “VIODE: A simulated dataset to address the challenges of visual–inertial odometry in dynamic environments,”RAL, 2021

    2021

  47. [47]

    4Seasons: A cross-season dataset for multi-weather SLAM in autonomous driving,

    P. Wenzelet al., “4Seasons: A cross-season dataset for multi-weather SLAM in autonomous driving,” inGCPR, 2020

    2020

  48. [48]

    Tartanair: A dataset to push the limits of visual slam,

    W. Wanget al., “Tartanair: A dataset to push the limits of visual slam,” inIROS, 2020

    2020

  49. [49]

    ORB-SLAM3: An accurate open-source library for visual, visual–inertial, and multimap SLAM,

    C. Camposet al., “ORB-SLAM3: An accurate open-source library for visual, visual–inertial, and multimap SLAM,”TRO, 2021

    2021

  50. [50]

    Dynavins: A visual-inertial slam for dynamic environ- ments,

    S. Songet al., “Dynavins: A visual-inertial slam for dynamic environ- ments,”RAL, 2022

    2022

  51. [51]

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

    D. Maggioet al., “Vggt-slam: Dense rgb slam optimized on the sl (4) manifold,”NeurIPS, 2026

    2026

  52. [52]

    A tutorial on quantitative trajectory evaluation for visual–inertial odometry,

    Z. Zhanget al., “A tutorial on quantitative trajectory evaluation for visual–inertial odometry,” inIROS, 2018

    2018

This paper was first reviewed by deepseek-v4-flash on August 1, 2026.