Pith. sign in

REVIEW 4 major objections 5 minor 29 references

Neuromorphic Visual Odometry System for Intelligent Vehicle Application with Bio-inspired Vision Sensor

T0 review · 4 major / 5 minor · reviewed 2026-08-14 · deepseek-v4-flash

Pith's one-line read The paper claims the first in-vehicle visual odometry system that uses only a neuromorphic camera, validated on a real urban driving dataset with average planar localization error 0.581 m (relative 0.5578%).

desk verdict A competent application study of event-camera VO on driving data, but the 'first event-only' claim is unresolved because the paper never states whether the borrowed Zhu et al. tracker received IMU data. read the letter →

arxiv 1909.02490 v1 pith:3BP7BSKA submitted 2019-09-05 cs.RO

classification cs.RO
keywords neuromorphicvisioneventcameravisualodometryfeaturetrackingdepthfilterbundleadjustmentintelligentvehiclesurbandrivingdataset
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

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

The reading

This paper tries to show that a car can estimate its own motion using only a neuromorphic (event) camera, without any regular frame camera or inertial sensor. It builds a visual odometry system that turns the asynchronous event stream into event frames, tracks feature points through those frames, and recovers camera pose by bundle adjustment and depth by triangulation with a probabilistic depth filter. The system is tested on an actual driving dataset over a 439 m urban sequence, where it reports an average planar localization error of 0.581 m and a relative position error of 0.5578%. The authors claim this is the first in-vehicle visual odometry system that uses only a neuromorphic camera, and the point of the claim is that event cameras' low latency, high dynamic range, and sparse output could make them viable for autonomous driving perception.

What carries the argument

The load-bearing mechanism is the event-frame feature-tracking pipeline from [14], used as the front end. Events are accumulated over a short time window into a frame; an Expectation-Maximization optical-flow step warps those events back to a common time, producing a corrected edge-like image; Harris corners are detected on that image; and features are matched between successive corrected frames by affine alignment. Around this front end, the system wraps a standard keyframe visual odometry loop: bundle adjustment for pose via PnP, triangulation of matched points, a depth filter for map-point insertion, and an eight-point-algorithm bootstrap. The front end is what converts the asynchronous event stream into the sparse, trackable features that the rest of the pipeline depends on.

What would settle it

Run the feature tracker and the full odometry pipeline on the same urban driving sequence twice: once with the IMU channel completely disabled and once with it enabled, comparing tracked feature lifetimes and final trajectory error. If disabling the IMU collapses the tracks or substantially increases the error, the paper's event-only claim is falsified.

Watch

Extended reading notes

Core claim

The central discovery is a feasibility result: an event-camera-only visual odometry pipeline can run on real road data and keep the vehicle on track to sub-meter accuracy, with 0.581 m average planar error and 0.5578% relative position error over 439 m. The system borrows the event-frame feature tracker from [14], aggregates events into frames, corrects them with EM optical flow, detects Harris corners, aligns features between warped frames, then estimates pose through bundle adjustment and builds a sparse map by triangulation filtered by a Gaussian plus uniform depth model. Bootstrapping uses the eight-point algorithm on the first two event frames because no map exists at startup. The experiments also show that tracked features live long enough on average (16.488 frames by day, 14.321 at night) for the depth filter to converge, and that tracking degrades only mildly at night compared with conventional cameras. The authors are explicit that the system currently runs offline only, because the feature tracker is too slow to track about 100 points in real time.

Load-bearing premise

The event-only claim assumes the feature-tracking code taken from [14] can produce stable tracks without any IMU data; the paper does not state whether the inertial channel of the driving dataset was supplied to that code.

Editorial extensions

If this is right

  • If the claimed error numbers hold, event-camera-only odometry is accurate enough for some automotive positioning uses on smooth urban roads, with relative position error below 0.6%.
  • The reported feature lifetimes of over 14 frames on average in both day and night indicate that event-based tracking is stable enough to feed a depth filter, not just for lab demonstrations.
  • The sharp error increase after the vehicle crosses two bumps shows that road roughness, not illumination, is the current limiting condition for event-only visual odometry.
  • Because the system currently runs offline, real-time in-vehicle use waits on a faster feature tracker or a lighter tracking scheme.
  • The modular parallel tracking-and-mapping structure means a faster front end can be slotted in without redesigning the pose and depth estimation threads.

Reading between the lines

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

  • The borrowed tracking algorithm was originally part of a visual-inertial odometry system; if the released code silently expects IMU measurements, the 'event-only' claim would need qualification, and this is testable but not resolved in the paper.
  • A natural next experiment is to compare the same pipeline with and without the inertial channel on the same driving sequence; the 0.5578% error would then separate the event-camera contribution from any hidden inertial support.
  • The bump-induced drift suggests that event-only vehicle odometry would benefit most from adding gyroscope or wheel-odometry measurements around vertical accelerations, rather than adding a full frame camera.
  • If feature tracking becomes real-time, the same architecture could plausibly transfer to the motorcycle and drone sequences in the same dataset, where the sensor's low latency matters most.
Share X Bluesky LinkedIn Reddit HN

Signed reviews

No signed human review yet.

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

4 major / 5 minor

Summary. The paper proposes a visual odometry system for vehicles that uses only a neuromorphic (event) camera, building on a feature detection and tracking algorithm originally developed by Zhu et al. [14]. The system aggregates events into frames, performs EM optical-flow correction, detects Harris corners, and then runs a parallel tracking-and-mapping pipeline: pose estimation by bundle adjustment/PnP, depth estimation by triangulation with a probabilistic depth filter, and initialization by the eight-point algorithm. Experiments are reported on the MVSEC urban driving dataset, with feature-lifetime statistics for day and night scenes and a trajectory evaluation over one 439 m interval (12 s to 72 s), giving an average planar error of 0.581 m and a relative position error of 0.5578%. The paper explicitly states that feature tracking is not real-time and that the system currently runs in offline mode.

Significance. If the central claim were fully established, this would be a useful early demonstration of event-camera-only visual odometry in a driving scenario, complementing prior event-based VO work that relies on IMU or conventional images. The paper uses standard, well-understood geometric estimation machinery (PnP, bundle adjustment, depth filters, eight-point initialization) and evaluates on a public dataset, which are strengths. However, the headline claim that the system 'only uses a neuromorphic camera' is not yet supported because the paper does not disclose whether the borrowed tracker from [14] received IMU data, and the quantitative evidence rests on a single sequence with no error bars and no same-data baseline. The significance is therefore conditional on resolving these points.

major comments (4)
  1. [Sections III.A and IV.C] The paper never states whether the feature tracker taken from Zhu et al. [14] was executed with IMU data from the MVSEC dataset. This matters because [14] is an event-based visual-inertial odometry algorithm, and Table 1 explicitly classifies Zhu et al. 2017 as 'fusion with IMU based on Kalman filter.' If IMU measurements were supplied to the tracker, the resulting feature tracks and the reported 0.5578% relative position error were not produced by a system that uses only a neuromorphic camera, contradicting the abstract's claim. The authors should state exactly what inputs the source code consumes and, if the claim is to stand, provide a run of the tracker without IMU data.
  2. [Section IV.D] The quantitative evaluation is based on a single sequence of about 439 m (12 s to 72 s of the MVSEC dataset) with LOAM-based localization as ground truth. There are no error bars, no repeated runs with different initializations or parameters, and no baseline comparison on the same sequence (e.g., a standard frame-based VO, EVO, or another event-based VO). The comparison in the text to EVO's 0.2% relative error is drawn from a different dataset and is therefore not a controlled comparison. As written, the experiment demonstrates that the pipeline runs on one sequence rather than verifying the feasibility claim in the abstract.
  3. [Section IV.C] The paper states that the system 'is far from meeting the requirement of tracking about 100 feature points synchronously' and 'can only run in offline mode,' with feature tracking results recorded to a txt file and later fed into the odometry pipeline. This means the evaluated system is not a real-time in-vehicle visual odometry system. The abstract's characterization should be qualified, or the authors should demonstrate at least a synchronous version of the pipeline, even at reduced feature count.
  4. [Section III.A] The description of the EM optical-flow correction is under-specified: the text does not clarify whether the correction uses only the event stream or also uses IMU priors as in the original algorithm of [14]. Figure 2 is reproduced from [14] without stating which components are reused and which are modified. This ambiguity is directly load-bearing for the event-only claim and should be resolved.
minor comments (5)
  1. [Section IV.C] There is a typo in the sentence 'the real-time performance of this feature tracking algorithm is quite pool'; it should be 'poor.' Also, the caption of Fig. 14 says 'event basd visual odometry system' and should say 'based.'
  2. [References] References [25] and [26] are the same Longuet-Higgins 1981 paper, with [26] containing a typo ('alorithm'). They should be merged, or the duplicate removed.
  3. [Table 1] The symbols '√' and 'X' are used inconsistently: for Zhu et al. [14] the 'Event camera only' column is marked 'X,' which is correct, but the same symbol is used in other columns to mean different things. The table should use a clearer legend.
  4. [Equation (1)] Equation (1) is missing spacing or a multiplication symbol between the scalar factor 1/Z_i and K*exp(ξ^)P_i, making the formula hard to read. The Jacobian in Eq. (2) is described as 'a 26 Jacobian matrix' and should be written as a 2×6 Jacobian.
  5. [General] The word 'transfer' is used in several places (e.g., 'rotation and transfer between the first two event frames'); the standard term in visual odometry is 'translation.'

Circularity Check

0 steps flagged · score 0.0 of 10

No circular derivation: the reported error is measured against an external ground truth, and the borrowed feature tracker creates an unresolved input assumption rather than a circular reduction.

full rationale

The paper reports the average planar error (0.581 m) and relative position error (0.5578%) as experimental statistics for a 439 m urban segment of the MVSEC dataset (Section IV.D, Table III), not as quantities derived from fitted parameters. The pose pipeline is a standard geometric chain: essential-matrix bootstrapping (Section III.E), PnP/bundle adjustment via Eq. (1)-(5), triangulation Eq. (6)-(7), and a Gaussian+Uniform depth filter Eq. (8). None of these equations encodes the reported error, and no parameter is fit to MVSEC and then re-labeled as a prediction. The feature tracker is imported from external prior work (Section III.A: 'The feature detection and feature tracking algorithm utilized in our system is proposed by Zhu et al. in [14]'; Section IV.C: 'Using the source code provided by Alex Zhu et al.'), so the paper is not self-citing a load-bearing theorem. The one substantive concern is that the headline 'only uses a neuromorphic camera' claim is left undetermined because the paper never states whether that borrowed source code consumed IMU data, and Table I classifies Zhu et al. [14] as 'fusion with IMU based on Kalman filter.' That is a missing-support/correctness issue, not a circular reduction: the reported numbers are measurements from an external dataset, not consequences of an assumed conclusion. No circular step can be exhibited, so the appropriate score is 0.

Assumptions & free parameters 4 free parameters · 4 assumptions · 0 invented entities

No new physical entities, forces, or conserved quantities are introduced; the 'ledger' here is about experimental and pipeline assumptions rather than theory postulates.

free parameters (4)
  • Depth filter convergence threshold
    Controls when a tracked feature is inserted into the global map; the paper says 'when the uncertainty of the depth distribution is smaller than the threshold' but never gives the threshold value.
  • Key frame interval
    System 'only regularly stores an event frame as a key frame'; the period is unspecified, though it affects feature detection frequency and depth filter resets.
  • Event aggregation time interval
    Event points in 'an adjustable time interval' are aggregated into event frames; the interval is not specified.
  • Feature lifetime cutoff = 3 frames
    Feature points with lifetime less than 3 event frames are ignored when computing lifetime statistics (Section IV.C).
assumptions (4)
  • standard math Essential matrix, PnP, and bundle adjustment constraints model camera motion from point correspondences (Section III.B, III.E).
    Standard multi-view geometry results invoked without proof.
  • domain assumption Depth measurements follow a Gaussian+Uniform mixture model with an inlier probability (Eq.8, from Vogiatzis and Hernandez [23]).
    Assumed distribution for depth filter updates; the paper adopts it from prior work without validation on event-camera data.
  • domain assumption The feature tracking algorithm from Zhu et al. [14] runs correctly on event frames without inertial data (Section III.A).
    The algorithm originates from a visual-inertial odometry paper; the present paper does not state whether IMU data was used when running the provided source code.
  • domain assumption LOAM output provides accurate ground truth for the vehicle trajectory (Section IV.D).
    LOAM itself is a state estimate with its own drift; treating it as ground truth for sub-meter error claims is an unvalidated assumption.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Neuromorphic Visual Odometry System for Intelligent Vehicle Application with Bio-inspired Vision Sensor." pith.science (2026). https://pith.science/paper/3BP7BSKA

@misc{pith2026190902490,
  author       = {Pith},
  title        = {Pith review of: Neuromorphic Visual Odometry System for Intelligent Vehicle Application with Bio-inspired Vision Sensor},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/3BP7BSKA}},
  note         = {Machine review of arXiv:1909.02490}
}
read the original abstract

The neuromorphic camera is a brand new vision sensor that has emerged in recent years. In contrast to the conventional frame-based camera, the neuromorphic camera only transmits local pixel-level changes at the time of its occurrence and provides an asynchronous event stream with low latency. It has the advantages of extremely low signal delay, low transmission bandwidth requirements, rich information of edges, high dynamic range etc., which make it a promising sensor in the application of in-vehicle visual odometry system. This paper proposes a neuromorphic in-vehicle visual odometry system using feature tracking algorithm. To the best of our knowledge, this is the first in-vehicle visual odometry system that only uses a neuromorphic camera, and its performance test is carried out on actual driving datasets. In addition, an in-depth analysis of the results of the experiment is provided. The work of this paper verifies the feasibility of in-vehicle visual odometry system using neuromorphic cameras.

Figures

Figures reproduced from arXiv: 1909.02490 by the authors.

Figure 1
Figure 1. The framework of our event based visual odometry system [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. Fig.2.(d) shows two continuous corrected event frames before [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗
Figure 3
Figure 3. Fig.3. And the detail of update of [PITH_FULL_IMAGE:figures/full_fig_p003_3.png] view at source ↗
Figures from the paper (10 more)
Figure 2
Figure 2. Figure 2: The procedure of feature tracking: (a) the ordinary event stream in spatiotemporal coordinate system; (b) events aggregated in a event frame without [PITH_FULL_IMAGE:figures/full_fig_p004_2.png]
Figure 3
Figure 3. Figure 3: Depth-filter update for a new measurement [PITH_FULL_IMAGE:figures/full_fig_p004_3.png]
Figure 5
Figure 5. Figure 5: The grayscale images of daytime in Multi Vehicle Stereo Event [PITH_FULL_IMAGE:figures/full_fig_p005_5.png]
Figure 4
Figure 4. Figure 4: Some scenarios in Event-Camera Dataset and Simulator dataset [PITH_FULL_IMAGE:figures/full_fig_p005_4.png]
Figure 10
Figure 10. Figure 10: The feature tracking in four continuous night event frames (the same [PITH_FULL_IMAGE:figures/full_fig_p006_10.png]
Figure 8
Figure 8. Figure 8: The grayscale images of night in Multi Vehicle Stereo Event Camera [PITH_FULL_IMAGE:figures/full_fig_p006_8.png]
Figure 11
Figure 11. Figure 11: The distribution of lifetime in daytime scenario [PITH_FULL_IMAGE:figures/full_fig_p006_11.png]
Figure 9
Figure 9. Figure 9: The event frames of night in Multi Vehicle Stereo Event Camera [PITH_FULL_IMAGE:figures/full_fig_p006_9.png]
Figure 14
Figure 14. Figure 14: The localization error of our event basd visual odometry system in [PITH_FULL_IMAGE:figures/full_fig_p007_14.png]
Figure 13
Figure 13. Figure 13: The localization estimation from our event based visual odometry [PITH_FULL_IMAGE:figures/full_fig_p007_13.png]

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

29 extracted references · 29 canonical work pages

  1. [14]

    Event-based visual inertial odometry,

    A. Z. Zhu, N. Atanasov, and K. Daniilidis, “Event-based visual inertial odometry,” in IEEE Conference on Computer Vision & Pattern Recog- nition, 2017

  2. [1]

    Neuromorphic vision based multivehicle detection and tracking for intelligent transportation system,

    G. Chen, H. Cao, M. Aafaque, J. Chen, and C. Ye, “Neuromorphic vision based multivehicle detection and tracking for intelligent transportation system,” Journal of Advanced Transportation , vol. 2018, no. 4815383, p. 13, 2018

  3. [2]

    Evo: A geometric approach to event-based 6-dof parallel tracking and mapping in real-time,

    H. Rebecq, T. Horstschaefer, G. Gallego, and D. Scaramuzza, “Evo: A geometric approach to event-based 6-dof parallel tracking and mapping in real-time,” IEEE Robotics & Automation Letters , vol. 2, no. 2, pp. 593–600, 2017

  4. [3]

    Interacting maps for fast visual interpretation,

    M. Cook, L. Gugelmann, F. Jug, C. Krautz, and A. Steger, “Interacting maps for fast visual interpretation,” in International Joint Conference on Neural Networks , 2011

  5. [4]

    Simul- taneous mosaicing and tracking with an event camera,

    H. Kim, A. Handa, R. Benosman, S.-H. Ieng, and A. Davison, “Simul- taneous mosaicing and tracking with an event camera,” in Proceedings of the British Machine Vision Conference . BMV A Press, 2014

  6. [5]

    Weikersdorfer, R

    D. Weikersdorfer, R. Hoffmann, and J. Conradt, Simultaneous Localiza- tion and Mapping for Event-Based Vision Systems , 2013

  7. [6]

    Low-latency event-based visual odome- try,

    A. Censi and D. Scaramuzza, “Low-latency event-based visual odome- try,” inIEEE International Conference on Robotics & Automation , 2014

  8. [7]

    Event- based 3d slam with a depth-augmented dynamic vision sensor,

    D. Weikersdorfer, D. B. Adrian, D. Cremers, and J. Conradt, “Event- based 3d slam with a depth-augmented dynamic vision sensor,” in IEEE International Conference on Robotics & Automation , 2014

Show all 29 references
  1. [8]

    Event-based, 6-dof pose tracking for high-speed maneuvers,

    E. Mueggler, B. Huber, and D. Scaramuzza, “Event-based, 6-dof pose tracking for high-speed maneuvers,” in IEEE/RSJ International Confer- ence on Intelligent Robots & Systems , 2014

  2. [9]

    Event-based, 6-dof camera tracking for high- speed applications,

    G. Gallego, J. E. A. Lund, E. Mueggler, H. Rebecq, T. Delbruck, and D. Scaramuzza, “Event-based, 6-dof camera tracking for high- speed applications,” IEEE Transactions on Pattern Analysis & Machine Intelligence, vol. PP, no. 99, pp. 1–1, 2016

  3. [10]

    Low-latency visual odometry using event-based feature tracks,

    B. Kueng, E. Mueggler, G. Gallego, and D. Scaramuzza, “Low-latency visual odometry using event-based feature tracks,” in IEEE/RSJ Inter- national Conference on Intelligent Robots & Systems , 2016

  4. [11]

    H. Kim, S. Leutenegger, and A. J. Davison, Real-Time 3D Reconstruc- tion and 6-DoF Tracking with an Event Camera , 2016

  5. [12]

    Emvs: Event- based multi-view stereo3d reconstruction with an event camera in real- time,

    H. Rebecq, G. Gallego, E. Mueggler, and D. Scaramuzza, “Emvs: Event- based multi-view stereo3d reconstruction with an event camera in real- time,” International Journal of Computer Vision , no. 2, pp. 1–21, 2017

  6. [13]

    Accurate angular velocity estimation with an event camera,

    G. Gallego and D. Scaramuzza, “Accurate angular velocity estimation with an event camera,” IEEE Robotics & Automation Letters , vol. PP, no. 99, pp. 1–1, 2017

  7. [15]

    Continuous-time visual-inertial odometry for event cameras,

    M. Elias, G. Guillermo, R. Henri, and S. Davide, “Continuous-time visual-inertial odometry for event cameras,” IEEE Transactions on Robotics, vol. PP, no. 99, pp. 1–16, 2017

  8. [16]

    Ultimate slam? combining events, images, and imu for robust visual slam in hdr and high speed scenarios,

    A. R. Vidal, H. Rebecq, T. Horstschaefer, and D. Scaramuzza, “Ultimate slam? combining events, images, and imu for robust visual slam in hdr and high speed scenarios,” IEEE Robotics & Automation Letters , vol. 3, no. 2, pp. 994–1001, 2018

  9. [17]

    Svo: Fast semi-direct monocular visual odometry,

    C. Forster, M. Pizzoli, and D. Scaramuzza, “Svo: Fast semi-direct monocular visual odometry,” in IEEE International Conference on Robotics & Automation , 2014

  10. [18]

    Parallel tracking and mapping for small ar workspaces,

    G. Klein and D. Murray, “Parallel tracking and mapping for small ar workspaces,” in IEEE & Acm International Symposium on Mixed & Augmented Reality , 2008

  11. [19]

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

    R. Murartal, 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, 2017

  12. [20]

    Engel, T

    J. Engel, T. Schps, and D. Cremers, LSD-SLAM: Large-Scale Direct Monocular SLAM , 2014

  13. [21]

    Harltey and A

    A. Harltey and A. Zisserman, Multiple view geometry in computer vision (2. ed.) , 2003

  14. [22]

    Computer vision and image under- standing,

    B. A. Maxwell and S. A. Shafer, “Computer vision and image under- standing,” Machine Learning & Data Mining Methods & Applications , vol. 72, no. 2, pp. 143–162(20), 2003

  15. [23]

    Video-based, real-time multi-view stereo ,

    G. V ogiatzis and C. Hernndez, “Video-based, real-time multi-view stereo ,” Image & Vision Computing , vol. 29, no. 7, pp. 434–441, 2011

  16. [24]

    A survey of planar homography estimation techniques,

    A. Agarwal, C. V . Jawahar, and P. J. Narayanan, “A survey of planar homography estimation techniques,” in Technical Reports, International Institute of Information Technology , 2005

  17. [25]

    A computer algorithm for reconstructing a scene from two projections,

    H. C. Longuet-Higgins, “A computer algorithm for reconstructing a scene from two projections,” Nature, vol. 293, no. 5828, pp. 133–135, 1981

  18. [26]

    A computer alorithm for reconstructing a scene from two projections,

    ——, “A computer alorithm for reconstructing a scene from two projections,” vol. 293, no. 5828, pp. 133–135, 1981

  19. [27]

    The event-camera dataset: Event-based data for pose estimation, vi- sual odometry, and slam,

    E. Mueggler, H. Rebecq, G. Gallego, T. Delbruck, and D. Scaramuzza, “The event-camera dataset: Event-based data for pose estimation, vi- sual odometry, and slam,” International Journal of Robotics Research , vol. 36, no. 49, pp. 142–149, 2017

  20. [28]

    The multi vehicle 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 multi vehicle stereo event camera dataset: An event camera dataset for 3d perception,” IEEE Robotics & Automation Letters , vol. 3, no. 3, pp. 2032–2039, 2018

  21. [29]

    Lifetime estimation of events from dynamic vision sensors,

    E. Mueggler, C. Forster, N. Baumli, G. Gallego, and D. Scaramuzza, “Lifetime estimation of events from dynamic vision sensors,” in IEEE International Conference on Robotics & Automation , 2015

Pith tools

Reviewed August 14, 2026 · model on record in the stance chip above.