Pith. sign in

REVIEW 3 major objections 6 minor 1 cited by

This paper claims that ego-motion can be estimated directly from sparse optical flow through a generalized 3D ray-based motion-field equation, making visual odometry fast and camera-model agnostic.

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 ↗

SMF-VO estimates a camera's velocity directly from sparse optical flow with a ray-based motion-field equation, reporting >100 FPS on a Raspberry Pi with competitive visual-odometry accuracy.

T0 review reviewed 2026-08-03 challenge →

load-bearing objection Fast, cleanly derived system, but the 'motion-centric, bypass pose estimation' claim collapses under its own ablation: the reported accuracy depends on the nonlinear pose-landmark optimization the paper calls optional. the 3 major comments →

arxiv 2511.09072 v2 pith:GEPRWZHC submitted 2025-11-12 cs.RO cs.CV

SMF-VO: Direct Ego-Motion Estimation via Sparse Motion Fields

Sangheon Yang , Yeongin Yoon , Hong Mo Jung , Jongwoo Lim This is my paper
classification cs.RO cs.CV
keywords visual odometryego-motion estimationmotion fieldsparse optical flowlinear velocityfisheye cameraembedded visionRANSAC
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

This paper proposes a visual odometry system that skips the usual pipeline of tracking landmarks, estimating camera poses, and optimizing a map. Instead, it uses the classical motion-field relationship—how image feature motion depends on camera motion—to solve for the camera's instantaneous linear and angular velocity in a small linear least-squares problem each frame. Because the formulation works on normalized 3D rays rather than 2D pixels, it also handles wide-angle and fisheye lenses without special-casing. The authors show that this motion-centric approach reaches roughly one-tenth of a meter trajectory accuracy on standard benchmarks while running faster than 100 frames per second on a low-power CPU, an order of magnitude faster than conventional feature-based VO.

Core claim

The central claim is that ego-motion can be recovered directly from sparse optical flow by solving a 6x6 linear system derived from a generalized 3D ray-based motion field equation. For a point seen along ray r with depth d, the ray velocity r-dot is a linear function of the camera's angular velocity omega and linear velocity v: r-dot = [r]x omega + (r r^T - I) v / d. Stacking these constraints over a few tracked points yields an over-determined linear system; a RANSAC loop rejects outliers, and an optional lightweight nonlinear refinement of only the current keyframe and its landmarks keeps drift small. The authors argue that this eliminates the need for explicit pose estimation, multi-fram

What carries the argument

The central object is the generalized 3D ray-based motion field equation, which expresses the temporal derivative of a normalized viewing ray r as a linear function of the camera's velocity (angular omega, linear v): r-dot = [r]x omega + (r r^T - I) v / d, where d is the Euclidean depth. Since the equation is linear in (omega, v), each tracked feature contributes constraints, and several features form a 3n x 6 linear system whose least-squares solution gives the instantaneous motion. The ray representation, instead of 2D pixel coordinates, is what makes the formulation camera-model agnostic; in the pinhole case it reduces to the classical pixel-based motion field.

Load-bearing premise

The method depends on having reliable metric depth for every tracked point, obtained from stereo disparity or frame-to-frame triangulation; the paper itself notes that short-baseline stereo depth can be noisy, particularly for distant features, so biased, sparse, or unavailable depth degrades the velocity estimate exactly where lightweight embedded VO is most needed.

What would settle it

Perturb the depths supplied to the linear system by a known bias and measure the resulting error in the estimated velocity; if the velocity error grows proportionally to the depth bias, the method's claimed accuracy in real scenes rests entirely on the quality of the stereo depth input. A field check would be to run the same pipeline on a long corridor with distant, low-texture features where stereo depth is noisy and compare trajectory error to a pose-centric baseline under identical conditions.

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

If this is right

  • If the method holds up, accurate velocity estimates become available at embedded-computing cost, opening up visual odometry on drones, AR/VR headsets, and wearables that cannot run bundle adjustment.
  • Because no pose or map is maintained except lightweight keyframe poses, the per-frame cost scales only with the number of tracked sparse features, not the map size.
  • The same motion formulation can be dropped into existing feature-tracking pipelines: any camera model, including fisheye, can be used without redesigning the motion solver.
  • The velocity estimate can be integrated over time to produce a trajectory, and the paper shows this accumulated trajectory reaches accuracy comparable to pose-centric systems on benchmark datasets.
  • The optional nonlinear optimization is separate; for easy sequences it can be turned off for even faster operation with modest accuracy loss, which the paper demonstrates in its ablation.

Where Pith is reading between the lines

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

  • If depth from stereo is the dominant limiting factor, the framework could be paired with learned monocular depth or an IMU to push accuracy in the low-texture regimes where short-baseline stereo fails—an extension the paper leaves implicit.
  • A direct velocity signal at over 100 Hz could feed control loops (drones, balancing robots) without integrating positions first, making the estimator useful in ways pose-centric VO is not.
  • The linear structure may allow closed-form uncertainty propagation and sensor-failure detection, because the condition number of the 6x6 normal matrix directly reflects how well the observed feature geometry constrains each velocity component.
  • The ray-based equation's resemblance to spherical-projection motion fields suggests a straightforward port to event cameras, whose output is already a form of sparse motion.
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 proposes SMF-VO, a stereo visual odometry system that estimates the camera's instantaneous linear and angular velocity by solving a linear system derived from a ray-based motion field equation using sparse optical flow and metric depths from stereo. A RANSAC scheme selects inliers, and a bundle-adjustment-style nonlinear optimization is applied at keyframes to refine motion and landmark positions. The authors evaluate on EuRoC, KITTI, and TUM-VI Room, reporting RMSE ATE values around 0.1 m and runtimes below 10 ms/frame on a Raspberry Pi 5, claiming a new 'motion-centric' paradigm that bypasses explicit pose estimation and landmark tracking.

Significance. If the results hold, the proposed method offers a compelling speed-accuracy tradeoff on embedded hardware, with the ray-based motion field providing a clean generalization across camera models. The derivation is mostly sound, and the benchmark comparison is thorough. However, the central claim that the method bypasses pose estimation is not supported because the reported accuracy depends on a keyframe-based BA step; the motion-field solver alone yields an order-of-magnitude higher error (Table IV). The paper's contribution is better described as a fast parametric front-end combined with a lightweight BA, not a pure motion-centric VO.

major comments (3)
  1. [Abstract; §III-D; §IV-B; Table IV] The abstract and introduction state that SMF-VO 'bypasses the need for explicit pose estimation or expensive landmark tracking.' However, the full system includes the nonlinear optimization in §III-D, which jointly optimizes the current keyframe pose [R|t]_k and landmark positions P_l via Eq. (13). The ablation (Table IV) shows that removing this step degrades ray-based RMSE ATE on EuRoC from 0.128 m to 1.202 m average and on TUM-VI Room from 0.082 m to 0.754 m. The benchmark results in Tables I–III all use 'w/ nonlinear optimization.' Consequently, the demonstrated accuracy is not attributable to the motion-field velocity estimator alone, and calling the optimization 'optional' is misleading. The paper should either reframe the contribution as a hybrid system or provide evidence that the motion-field estimator alone reaches competitive accuracy.
  2. [§III-A, Eq. (4); Eq. (5)] Equation (4) for \dot{p}_y appears to have incorrect signs on the angular-velocity terms. The standard motion-field formula (e.g., Trucco & Verri) is \dot{p}_y = (v_z y - v_y f)/Z + \omega_x(f + y^2/f) - \omega_y(xy/f) - \omega_z x, whereas Eq. (4) gives -\omega_x f + \omega_z x + \omega_y xy/f - \omega_x y^2/f, which is the negative. The matrix A(p) in Eq. (5) matches the standard formula, so Eq. (4) and A(p) are inconsistent. Please correct the typo and verify that the pixel-based experiments used the correct equation.
  3. [§III-B, Eq. (11); §III-D] Equation (11) uses the Euclidean depth d as coefficients in the linear system. Section III-D acknowledges that short-baseline stereo depths 'can be noisy, particularly for distant features,' yet no experiment quantifies the sensitivity of the velocity estimate to depth noise. Since the claimed motion-centric contribution rests on a lightweight linear solve, it is important to show how depth errors affect accuracy, e.g., by perturbing depths or comparing stereo disparity depths vs. triangulated depths in the ablation.
minor comments (6)
  1. [Tables I–III] Report single-run RMSE ATE and timing without standard deviations; error bars or multiple runs would strengthen the comparison.
  2. [Tables I, II] V203 and KITTI 01 are excluded with qualitative justification; please provide quantitative failure statistics for these sequences to justify the exclusion.
  3. [Abstract] 'Over 100 FPS' is achieved on EuRoC and TUM-VI, but on KITTI the average is about 52 FPS (Table II). Please qualify this claim to avoid overgeneralization.
  4. [V-A, Experiments] The timing comparisons are not fully apples-to-apples: ROVIO reports only filter update time, and ORB-SLAM3 tracking time excludes preprocessing. This should be stated more prominently and, if possible, equalized.
  5. [General] No code or data are released, which limits reproducibility. The paper would benefit from a public implementation or at least a detailed parameter configuration.
  6. [Figures 3, 4] Captions contain duplicate '(c)' labels; renumber the subfigures.

Circularity Check

0 steps flagged

Derivation is self-contained; no circularity found.

full rationale

The motion-field equations (5)/(6) and (10)/(11) are derived in-paper from rigid-body kinematics (Eq. 2) and projection geometry; the ray-based result is independently checked against the spherical-projection motion field of ref. [29]. No fitted parameter is later renamed as a prediction: depths are inputs from stereo/triangulation, and the least-squares velocity solves are direct inversions of the motion-field constraints. The nonlinear 'Motion and Landmark Optimization' (Sec. III-D, Eq. 13) is a standard refinement stage rather than a quantity equivalent by construction to the velocity estimate; the ablation result that accuracy degrades without it is an empirical support issue, not a circularity. There are no load-bearing self-citations (references to the authors' own prior work are absent), no imported uniqueness theorem, and no ansatz hidden in a citation. The central derivation is therefore self-contained; the mismatch between the 'bypasses explicit pose estimation' abstract claim and the presence of Eq. (13) in the evaluated system is a correctness/consistency concern outside the circularity pass.

Axiom & Free-Parameter Ledger

4 free parameters · 4 axioms · 0 invented entities

The central derivation uses standard rigid-body kinematics and known projection geometry; no new physical entities are introduced. The main load-bearing assumptions are metric stereo depth and known camera calibration. Several hyperparameters are hand-selected without sensitivity analysis, but they do not constitute new theoretical constructs.

free parameters (4)
  • RANSAC parameters (Q, n_s, gamma_0, tau_pi, tau_u) = Q=99.99%, n_s=3, gamma_0=0.9; tau_pi/tau_u not numerically specified
    Hand-selected in Algorithm 1; no sensitivity analysis is provided, and these thresholds control the inlier set and thus the final velocity estimate.
  • Cauchy robust-loss scale c
    Called a 'tunable scale parameter' in Section III-D; no default value or sensitivity study is given, but it affects the nonlinear refinement that dramatically improves accuracy.
  • Keyframe selection thresholds (tau_n, time, relative motion)
    Defined in Section III-D as 'standard practices' but no numerical values or criteria are reported; these determine how often the nonlinear pose-landmark optimization runs.
  • KLT feature tracking parameters (number of features, window size, etc.)
    The system relies on KLT sparse tracking but the paper does not specify feature count, patch size, or pyramid settings, which materially affect runtime and accuracy.
axioms (4)
  • domain assumption Rigid-body motion model: dot(P) = -v - omega x P (Eq. 2).
    Assumes the world is static and the camera moves as a rigid body; independently moving objects must be rejected by RANSAC.
  • domain assumption Accurate metric depth for each tracked point is available from stereo or triangulation (Eqs. 6 and 11).
    The velocity estimate is linearly dependent on inverse depth; biased or noisy stereo depth directly biases the pose-integrated trajectory.
  • domain assumption KLT sparse optical flow approximates the true motion field between consecutive frames.
    The method computes ray velocities from KLT tracks; photometric changes, large motions, or occlusions violate this assumption, though RANSAC mitigates outliers.
  • domain assumption The camera projection/unprojection model is known and calibrated, including for fisheye lenses.
    The ray-based formulation requires normalized rays as input; the paper does not specify the unprojection model used for TUM-VI fisheye images.

reviewed 2026-08-03 · how reviews work

0 comments
Cite this review

Pith. "Pith review of SMF-VO: Direct Ego-Motion Estimation via Sparse Motion Fields." pith.science (2026). https://pith.science/paper/GEPRWZHC

@misc{pith2026251109072,
  author       = {Pith},
  title        = {Pith review of: SMF-VO: Direct Ego-Motion Estimation via Sparse Motion Fields},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/GEPRWZHC}},
  note         = {Machine review of arXiv:2511.09072}
}
Share X Bluesky LinkedIn Reddit HN
read the original abstract

Traditional Visual Odometry (VO) and Visual Inertial Odometry (VIO) methods rely on a 'pose-centric' paradigm, which computes absolute camera poses from the local map thus requires large-scale landmark maintenance and continuous map optimization. This approach is computationally expensive, limiting their real-time performance on resource-constrained devices. To overcome these limitations, we introduce Sparse Motion Field Visual Odometry (SMF-VO), a lightweight, 'motion-centric' framework. Our approach directly estimates instantaneous linear and angular velocity from sparse optical flow, bypassing the need for explicit pose estimation or expensive landmark tracking. We also employed a generalized 3D ray-based motion field formulation that works accurately with various camera models, including wide-field-of-view lenses. SMF-VO demonstrates superior efficiency and competitive accuracy on benchmark datasets, achieving over 100 FPS on a Raspberry Pi 5 using only a CPU. Our work establishes a scalable and efficient alternative to conventional methods, making it highly suitable for mobile robotics and wearable devices.

Figures

Figures reproduced from arXiv: 2511.09072 by Hong Mo Jung, Jongwoo Lim, Sangheon Yang, Yeongin Yoon.

Figure 1
Figure 1. Figure 1: An overall comparison of the accuracy and speed of [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Overview of the framework we propose. A. 3D Camera Motion from Sparse Motion Field For a 3D point in the camera coordinate frame, P = [X Y Z] ⊤, the corresponding point projected onto the image plane at focal length f can be expressed as p = f (P/Z) := [x y f] ⊤. (1) When the camera moves with the linear velocity v and angular velocity ω, the movement (velocity) of the point P is described by P˙ = −v − ω ×… view at source ↗
Figure 3
Figure 3. Figure 3: Qualitative trajectory result of EuRoC sequences. The estimated trajectories of SMF-VO (ours), ORB-SLAM3 [28] [PITH_FULL_IMAGE:figures/full_fig_p006_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: Qualitative trajectory result of KITTI sequences. The estimated trajectories of SMF-VO (ours), ORB-SLAM3 [28] [PITH_FULL_IMAGE:figures/full_fig_p006_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: Qualitative trajectory results for TUM-VI Room se [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.

Forward citations

Cited by 1 Pith paper

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

  1. LC-Flow: Learning Local Continuous Optical Flow and Confidence from events

    cs.CV 2026-05 unverdicted novelty 7.0

    LC-Flow introduces a continuous local recurrent network for learning sparse optical flow and confidence directly from event streams, with confidence-guided aggregation reaching new SOTA on MVSEC.

Reference graph

Works this paper leans on

39 extracted references · 1 linked inside Pith · cited by 1 Pith paper

  1. [1]

    Real-time simultaneous localisation and mapping with a single camera,

    A. J. Davison, “Real-time simultaneous localisation and mapping with a single camera,” inProceedings of the Ninth IEEE International Conference on Computer Vision (ICCV), vol. 2. IEEE, October 2003, pp. 1403–1410

    2003

  2. [2]

    Real-time localization and 3d reconstruction,

    E. Mouragnon, M. Lhuillier, M. Dhome, F. Dekeyser, and P. Sayd, “Real-time localization and 3d reconstruction,” inProceedings of the 2006 IEEE Computer Society Conference on Computer Vision and Pattern Recognition (CVPR’06), vol. 1. IEEE, June 2006, pp. 363– 370

    2006

  3. [3]

    Parallel tracking and mapping for small ar workspaces,

    G. Klein and D. Murray, “Parallel tracking and mapping for small ar workspaces,” inProceedings of the 6th IEEE and ACM International Symposium on Mixed and Augmented Reality (ISMAR). IEEE, November 2007, pp. 225–234

    2007

  4. [4]

    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, no. 5, pp. 1147–1163, 2015

    2015

  5. [5]

    Stereoscan: Dense 3d recon- struction in real-time,

    A. Geiger, J. Ziegler, and C. Stiller, “Stereoscan: Dense 3d recon- struction in real-time,” inProceedings of the 2011 IEEE Intelligent V ehicles Symposium (IV). IEEE, June 2011, pp. 963–968

    2011

  6. [6]

    Orb-slam2: An open-source slam system for monocular, stereo, and rgb-d cameras,

    R. Mur-Artal and J. D. Tard ´os, “Orb-slam2: An open-source slam system for monocular, stereo, and rgb-d cameras,”IEEE Transactions on Robotics, vol. 33, no. 5, pp. 1255–1262, 2017

    2017

  7. [7]

    Lsd-slam: Large-scale direct monocular slam,

    J. Engel, T. Sch ¨ops, and D. Cremers, “Lsd-slam: Large-scale direct monocular slam,” inECCV. Springer International Publishing, September 2014, pp. 834–849

    2014

  8. [8]

    Svo: Semidirect visual odometry for monocular and multicamera systems,

    C. Forster, Z. Zhang, M. Gassner, M. Werlberger, and D. Scaramuzza, “Svo: Semidirect visual odometry for monocular and multicamera systems,”IEEE Transactions on Robotics, vol. 33, no. 2, pp. 249– 265, 2016

    2016

  9. [9]

    Direct sparse odometry,

    J. Engel, V . Koltun, and D. Cremers, “Direct sparse odometry,”IEEE TPAMI, vol. 40, no. 3, pp. 611–625, 2017

    2017

  10. [10]

    Double window optimisation for constant time visual slam,

    H. Strasdat, A. J. Davison, J. M. Montiel, and K. Konolige, “Double window optimisation for constant time visual slam,” inProceedings of the 2011 International Conference on Computer Vision (ICCV). IEEE, November 2011, pp. 2352–2359

    2011

  11. [11]

    g2o: A general framework for graph optimization,

    R. K ¨ummerle, G. Grisetti, H. Strasdat, K. Konolige, and W. Burgard, “g2o: A general framework for graph optimization,” inProceedings of the 2011 IEEE International Conference on Robotics and Automation (ICRA). IEEE, May 2011, pp. 3607–3613

    2011

  12. [12]

    A multi-state constraint kalman filter for vision-aided inertial navigation,

    A. I. Mourikis and S. I. Roumeliotis, “A multi-state constraint kalman filter for vision-aided inertial navigation,” inIEEE International Con- ference on Robotics and Automation. IEEE, April 2007, pp. 3565– 3572

    2007

  13. [13]

    Iterated extended kalman filter based visual-inertial odometry using direct pho- tometric feedback,

    M. Bloesch, M. Burri, S. Omari, M. Hutter, and R. Siegwart, “Iterated extended kalman filter based visual-inertial odometry using direct pho- tometric feedback,”The International Journal of Robotics Research, vol. 36, no. 10, pp. 1053–1072, 2017

    2017

  14. [14]

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

    Y . Fan, T. Zhao, and G. Wang, “Schurvins: Schur complement-based lightweight visual inertial navigation system,” inProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), June 2024, pp. 17 964–17 973

    2024

  15. [15]

    Vins-mono: A robust and versatile monoc- ular visual-inertial state estimator,

    T. Qin, P. Li, and S. Shen, “Vins-mono: A robust and versatile monoc- ular visual-inertial state estimator,”IEEE Transactions on Robotics, vol. 34, no. 4, pp. 1004–1020, 2018

    2018

  16. [16]

    Visual-inertial mapping with non-linear factor recovery,

    V . Usenko, N. Demmel, D. Schubert, J. St ¨uckler, and D. Cre- mers, “Visual-inertial mapping with non-linear factor recovery,”IEEE Robotics and Automation Letters, vol. 5, no. 2, pp. 422–429, 2019

    2019

  17. [17]

    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. [Online]. Available: https://arxiv.org/abs/2202.09199

    Pith/arXiv arXiv 2022

  18. [18]

    The interpretation of a moving retinal image,

    H. C. Longuet-Higgins and K. Prazdny, “The interpretation of a moving retinal image,”Proceedings of the Royal Society of London. Series B. Biological Sciences, vol. 208, no. 1173, pp. 385–397,

  19. [19]

    Trucco and A

    E. Trucco and A. Verri,Introductory Techniques for 3-D Computer Vision. Englewood Cliffs: Prentice Hall, 1998, vol. 201

    1998

  20. [20]

    The multivehicle stereo event camera dataset: An event camera dataset for 3d perception,

    A. Z. Zhu, D. Thakur, T. ¨Ozaslan, B. Pfrommer, V . Kumar, and K. Daniilidis, “The multivehicle stereo event camera dataset: An event camera dataset for 3d perception,”IEEE Robotics and Automation Letters, vol. 3, no. 3, pp. 2032–2039, 2018

    2032

  21. [21]

    Ev-flownet: Self-supervised optical flow estimation for event-based cameras,

    A. Z. Zhu, L. Yuan, K. Chaney, and K. Daniilidis, “Ev-flownet: Self-supervised optical flow estimation for event-based cameras,” in Proceedings of Robotics: Science and Systems (RSS), Pittsburgh, Pennsylvania, June 2018

    2018

  22. [22]

    Dsec: A stereo event camera dataset for driving scenarios,

    M. Gehrig, W. Aarents, D. Gehrig, and D. Scaramuzza, “Dsec: A stereo event camera dataset for driving scenarios,”IEEE Robotics and Automation Letters, vol. 6, no. 3, pp. 4947–4954, 2021

    2021

  23. [23]

    E-raft: Dense optical flow from event cameras,

    M. Gehrig, M. Millh ¨ausler, D. Gehrig, and D. Scaramuzza, “E-raft: Dense optical flow from event cameras,” inProceedings of the 2021 International Conference on 3D Vision (3DV). IEEE, December 2021, pp. 197–206

    2021

  24. [24]

    Unsupervised event-based learning of optical flow, depth, and egomotion,

    A. Z. Zhu, L. Yuan, K. Chaney, and K. Daniilidis, “Unsupervised event-based learning of optical flow, depth, and egomotion,” inCVPR. IEEE, 2019, pp. 989–997

    2019

  25. [25]

    Unsupervised learning of dense optical flow, depth and egomotion with event-based sensors,

    C. Ye, A. Mitrokhin, C. Ferm ¨uller, J. A. Yorke, and Y . Aloimonos, “Unsupervised learning of dense optical flow, depth and egomotion with event-based sensors,” inIEEE/RSJ International Conference on Intelligent Robots and Systems. IEEE, October 2020, pp. 5831–5838

    2020

  26. [26]

    Secrets of event- based optical flow, depth and ego-motion estimation by contrast maximization,

    S. Shiba, Y . Klose, Y . Aoki, and G. Gallego, “Secrets of event- based optical flow, depth and ego-motion estimation by contrast maximization,”IEEE TPAMI, 2024

    2024

  27. [27]

    Event-based visual inertial velometer,

    X. Lu, Y . Zhou, J. Niu, S. Zhong, and S. Shen, “Event-based visual inertial velometer,” inProceedings of Robotics: Science and Systems (RSS). Delft, Netherlands: CKS Robotic Institute, Hong Kong University of Science and Technology and School of Robotics, Hunan University, July 2024, equal contribution by Xiuyuan Lu and Yi Zhou

    2024

  28. [28]

    Orb-slam3: An accurate open-source library for visual, vi- sual–inertial, and multimap slam,

    C. Campos, R. Elvira, J. J. G. Rodr ´ıguez, J. M. Montiel, and J. D. Tard´os, “Orb-slam3: An accurate open-source library for visual, vi- sual–inertial, and multimap slam,”IEEE Transactions on Robotics, vol. 37, no. 6, pp. 1874–1890, 2021

    2021

  29. [29]

    Determining motion directly from normal flows upon the use of a spherical eye platform,

    T.-W. Hui and R. Chung, “Determining motion directly from normal flows upon the use of a spherical eye platform,” inProceedings of the IEEE Conference on Computer Vision and Pattern Recognition, 2013, pp. 2267–2274

    2013

  30. [30]

    Hartley and A

    R. Hartley and A. Zisserman,Multiple View Geometry in Computer Vision. Cambridge University Press, 2003

    2003

  31. [31]

    Complete solution classification for the perspective-three-point problem,

    X.-S. Gao, X.-R. Hou, J. Tang, and H.-F. Cheng, “Complete solution classification for the perspective-three-point problem,”IEEE TPAMI, vol. 25, no. 8, pp. 930–943, August 2003

    2003

  32. [32]

    Random sample consensus: A paradigm for model fitting with applications to image analysis and automated cartography,

    M. A. Fischler and R. C. Bolles, “Random sample consensus: A paradigm for model fitting with applications to image analysis and automated cartography,”Communications of the ACM, vol. 24, no. 6, pp. 381–395, 1981

    1981

  33. [33]

    Lucas-kanade 20 years on: A unifying framework,

    S. Baker and I. Matthews, “Lucas-kanade 20 years on: A unifying framework,”International Journal of Computer Vision, vol. 56, pp. 221–255, 2004

    2004

  34. [34]

    Direct sparse visual- inertial odometry using dynamic marginalization,

    L. V . Stumberg, V . Usenko, and D. Cremers, “Direct sparse visual- inertial odometry using dynamic marginalization,” inIEEE Interna- tional Conference on Robotics and Automation. IEEE, May 2018, pp. 2510–2517

    2018

  35. [35]

    Openvins: A research platform for visual-inertial estimation,

    P. Geneva, K. Eckenhoff, W. Lee, Y . Yang, and G. Huang, “Openvins: A research platform for visual-inertial estimation,” inIEEE Interna- tional Conference on Robotics and Automation. IEEE, May 2020, pp. 4666–4672

    2020

  36. [36]

    The euroc micro aerial vehicle datasets,

    M. Burri, J. Nikolic, P. Gohl, T. Schneider, J. Rehder, S. Omari, M. W. Achtelik, and R. Siegwart, “The euroc micro aerial vehicle datasets,” The International Journal of Robotics Research, vol. 35, no. 10, pp. 1157–1163, 2016

    2016

  37. [37]

    Are we ready for autonomous driving? the kitti vision benchmark suite,

    A. Geiger, P. Lenz, and R. Urtasun, “Are we ready for autonomous driving? the kitti vision benchmark suite,” inCVPR. Providence, RI, USA: IEEE, June 2012, pp. 3354–3361

    2012

  38. [38]

    The tum vi benchmark for evaluating visual-inertial odometry,

    D. Schubert, T. Goll, N. Demmel, V . Usenko, J. St ¨uckler, and D. Cre- mers, “The tum vi benchmark for evaluating visual-inertial odometry,” inIEEE/RSJ International Conference on Intelligent Robots and Systems. Madrid, Spain: IEEE, October 2018, pp. 1680–1687

    2018

  39. [1980]

    Available: https://royalsocietypublishing.org/doi/abs/ 10.1098/rspb.1980.0057

    [Online]. Available: https://royalsocietypublishing.org/doi/abs/ 10.1098/rspb.1980.0057

    arXiv 1980

This paper was first reviewed by deepseek-v4-flash on August 3, 2026.