Pith. sign in

REVIEW 3 major objections 5 minor 20 references

Towards a physically realistic computationally efficient DVS pixel model

T0 review · 3 major / 5 minor · reviewed 2026-08-15 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗ Review JSON

Pith's one-line read A circuit-derived DVS pixel model predicts signal and noise while allowing 1000x longer timesteps.

desk verdict First-passage-time event generation is a genuinely useful new mechanism for DVS simulation, but the 1000x speedup claim is only verified in a first-order noise regime, not on the full circuit model or against real event data. read the letter →

arxiv 2505.07386 v1 pith:VONSLQMH submitted 2025-05-12 eess.IV

Rui Graca , Tobi Delbruck This is my paper · ORCID
classification eess.IV
keywords DynamicVisionSensorDVSpixelmodeleventcamerasimulationstochasticgenerationfirst-passagetimeOrnstein-Uhlenbecknoisecircuitanalysishighrange
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

Dynamic Vision Sensor (DVS) event cameras record brightness changes as asynchronous events, but existing pixel models either run fast enough for scene simulation while ignoring noise and bias dependence, or model the physics accurately only with impractically small timesteps. This paper tries to get both: a pixel model built from circuit-analysis transfer functions whose parameters are physically meaningful, plus a stochastic event generator that uses first-passage-time theory to estimate the probability that noise crosses the event threshold between simulation timesteps. If the model is right, it predicts both signal and noise across different illuminances and bias settings, and it allows timesteps about 1000 times longer than prior noise-accurate approaches, with wall-clock speedups above 100x. That would make offline bias optimization and realistic high-dynamic-range event dataset generation practical.

What carries the argument

The load-bearing machinery is a set of four transfer functions derived from the photoreceptor and source-follower buffer circuit—two second-order functions $Z_m(s)$ and $Z_{out}(s)$ relating photocurrent and photoreceptor current to node voltages, plus two first-order buffer functions $A_{sf}(s)$ and $Z_{outsf}(s)$—whose small-signal parameters are recomputed at every timestep so the model captures large-signal changes. On top of this, the stochastic event generator treats shot noise as a first-order Ornstein-Uhlenbeck process and, at each timestep, uses first-passage-time theory to compute the probability that the noise drove $V_{diff}$ across the ON or OFF threshold between timesteps, sampling the crossing time from the first-passage-time distribution. This between-timestep probability is what allows the simulation to use timesteps orders of magnitude longer than the noise cutoff frequency without missing noise-triggered events.

What would settle it

Run the model on a real DVS pixel in a regime where the photoreceptor current is not large compared with the source-follower current (or where no dominant pole separates $\tau_{pd}$ and $\tau_{sf}$), fix the threshold at a few $\sigma$, and compare the noise event rate predicted with a timestep near $0.5/f_c$ against the rate predicted with a timestep near $10^{-5}/f_c$; if the two rates differ significantly, the stochastic generator's core assumption fails. A simpler experimental version: measure event rates on a physical pixel under dim illumination and check whether the model's event-rate-versus-threshold curve still matches at long timesteps.

Watch

Extended reading notes

Core claim

The paper's central claim is that a DVS pixel can be simulated accurately and efficiently by separating the deterministic and stochastic parts of the response: large-signal differential equations derived from circuit analysis update the pixel's operating point each timestep, while an Ornstein-Uhlenbeck noise model plus first-passage-time crossing probabilities accounts for the random threshold crossings that occur in the intervals between timesteps. The authors show that the fitted transfer functions match measured noise power spectral densities at several illuminances and two bias settings, that the model reproduces the asymmetric slow-falling-edge behavior seen in real pixels, and that stochastic event generation keeps predicted noise event rates accurate even when the timestep is 0.5/fc, whereas naive threshold checking at the same timestep badly underestimates event rates. They conclude that simulation timesteps can be increased by three orders of magnitude without losing noise accuracy, and that fitting physically meaningful circuit parameters lets the model generalize over illuminance and bias.

Load-bearing premise

The load-bearing premise is that the pixel's shot noise behaves like a first-order Ornstein-Uhlenbeck process, which the authors state requires the photoreceptor current to be large compared with the source-follower current and a single dominant time constant to exist between the photoreceptor and source-follower poles; if those conditions fail, the between-timestep crossing probabilities are only approximate and the 1000x timestep advantage may not preserve accurate event rates.

Editorial extensions

If this is right

  • Bias settings for DVS cameras could be optimized in simulation before hardware is touched, since the same fitted parameters predict noise and signal at different illuminances and biases.
  • Event-camera training datasets for high-dynamic-range and low-light scenes could be generated from frame-based videos or scene descriptions with physically plausible noise events rather than idealized threshold crossings.
  • Array-level simulations can run more than 100x faster in wall-clock time than noise-accurate models with small timesteps, making full-scene physical simulation practical.
  • The model predicts the asymmetric response of real pixels, including the slow OFF-event trail after a brightness decrease, so synthetic event streams should better match the temporal statistics of real DVS output.
  • Because the parameters are physically meaningful (capacitances, currents, Early voltages), the model could be transferred to different pixel designs by fitting to their circuit simulation or measured data.

Reading between the lines

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

  • If the model holds up, the same first-passage-time trick could be applied to other noise-driven threshold-crossing sensors, such as spiking neurons or comparators in mixed-signal circuits, wherever the noise is roughly first-order low-pass.
  • The paper's stated plan to incorporate the model into an array-level simulator suggests a direct test: generated full-scene event streams could be evaluated on downstream tasks like optical flow or object detection, where the value of realistic noise can be quantified.
  • A natural extension would be to include flicker noise in the stochastic crossing calculation rather than only in the PSD fit; the paper notes flicker is small in practice, but in low-frequency event statistics it may still matter for long simulations.
  • The assumption of a dominant pole between the photoreceptor and source-follower time constants may fail at extreme bias settings; an explicit test of the model at such settings would reveal how much accuracy the 1000x timestep claim retains outside the fitted regime.
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

3 major / 5 minor

Summary. The paper proposes a DVS pixel model that combines circuit-analysis-derived large-signal differential equations with a stochastic event generation mechanism. The model consists of a second-order photoreceptor transfer function and a first-order source-follower buffer transfer function, with small-signal parameters updated at each timestep. Model parameters are physically meaningful circuit quantities (capacitances, bias currents, transistor parameters), fitted to measured noise PSDs from a DAVIS346 test pixel and to SPICE simulations. The key efficiency innovation is a stochastic event generator based on first-passage-time theory for an Ornstein-Uhlenbeck process, which computes the probability of a threshold crossing between simulation timesteps. The paper reports that this allows timesteps 1000x longer than previous methods without loss of noise event-rate accuracy, with validation in Fig. 6 performed on synthetic first-order low-pass noise. Large-signal step and pulse responses are shown in Figs. 4 and 5.

Significance. If the central claims hold, the model would be a substantial practical advance: it would enable array-level DVS simulation with physically meaningful dependence on illuminance and bias settings, supporting offline bias optimization and generation of realistic HDR event datasets. The paper's strengths include the physically interpretable parameter set, the good agreement of the fitted PSD model with pixel measurements across illuminances and biases (Fig. 3), and the explicit treatment of sub-timestep threshold crossings rather than simple sampled-threshold checks. However, the headline efficiency claim is currently conditional on an Ornstein-Uhlenbeck assumption whose validity conditions are stated but not verified for the fitted parameters, and the stochastic-generator validation in Fig. 6 is against the same first-order noise model used to construct the generator, not against the full second-order model or measured event streams. The physically realistic claims are therefore not yet fully supported by the evidence presented.

major comments (3)
  1. [Section V] The stochastic event generator is derived for first-order low-pass filtered white noise, and the text states this is valid when 'Ipr is large compared to Isf' and when 'there is a clear dominant pole between tau_pd and tau_sf'. The paper never reports the fitted values of Ipr, Isf, tau_pd, and tau_sf, nor the pole locations for the parameter set used in Figs. 3-6, and it does not test the crossing-probability formula against the full second-order model. Because the 1000x timestep claim rests directly on this assumption, please verify the stated conditions for the fitted pixel or provide simulation evidence that the full model's event rates are reproduced by the OU-based generator.
  2. [Fig. 6] The validation of the stochastic event generator is performed on synthetic first-order low-pass noise, not on the paper's second-order photoreceptor/buffer model and not on measured DVS event streams. The plateau in Fig. 6B demonstrates only that the Bernoulli/first-passage sampler reproduces OU crossing rates. To support the conclusion that the 1000x speedup preserves accurate noise event rates for the physical model, add a comparison against a low-timestep simulation of the full second-order model, and ideally against measured DVS noise event rates at the same illuminance and bias settings.
  3. [Figs. 4 and 5] The large-signal validation is qualitative. Fig. 4 compares simulated voltages and PSDs to the same fitted PSD model, which is at least partially circular, and Fig. 5 shows the 'asymmetrical non-linear behavior observed in practice' without a measured overlay or any error metric. Please include a quantitative comparison to measured pixel responses (for example, measured step or pulse transients or event-rate traces), or explicitly state that the large-signal validation is illustrative rather than quantitative.
minor comments (5)
  1. [Fig. 4 caption] The caption says '(E) and (F) show the PSD from (D) and (E)', which appears to be a typo; it should presumably refer to panels (C) and (D).
  2. [Abstract and Fig. 6C] The abstract claims timesteps 'greater than 1000x longer', while Fig. 6C reports wall-clock speedups of 'over 100x'. These are different metrics; please clarify in the text that the 1000x refers to timestep size and that wall-clock speedup is additionally reported.
  3. [Reference [1]] The DOI for reference [1] is given as '0.1109/JSSC.2007.914337'; the leading '1' appears to be missing and should be '10.1109/JSSC.2007.914337'.
  4. [Fig. 3] The label 'Some minor phenomena not yet modeled' is vague; specify which phenomena are omitted (for example, reset noise or parasitic coupling) and estimate their practical impact.
  5. [Section V] Please clarify how the time-varying signal level is incorporated into the first-passage-time crossing probability during large-signal transients, since the OU process description assumes a stationary mean and the simulation in Fig. 5 includes large deterministic changes in Ipd.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the stochastic event generator is validated against a fine-timestep simulation of the same explicit OU noise model, and the event-rate prediction is not a direct fit to event-rate data.

full rationale

The derivation chain is not circular. The circuit-derived transfer functions in Section II are quoted explicitly in the paper, with [19] cited for the derivation; they are checkable equations and are not identical to the paper's target predictions. The stochastic event generator in Section V is based on external first-passage-time results [20], [21], and Fig. 6 validates it by comparing large-timestep Bernoulli/first-passage event rates against fine-timestep direct simulation of the same OU noise model. That is a numerical consistency check of an approximation scheme, not a reduction of the output to the input. The fitted parameters are fitted to PSD and signal measurements, while event rates are a nontrivial function of those parameters rather than a direct fit to event-rate data. Self-citations [17]-[19],[22] supply circuit equations and methodology, but the load-bearing math is stated in the paper and the external [20],[21] method is independent. The only caveat is a missing-support issue, not circularity: Section V states that the OU assumption holds when Ipr is large compared to Isf and there is a clear dominant pole between tau_pd and tau_sf, but the paper does not verify these conditions for the fitted parameter sets, so the physical-event-rate half of the 1000x claim is conditional on that unverified regime. This limits the strength of the physical realism claim, but it does not make the derivation circular.

Assumptions & free parameters 4 free parameters · 5 assumptions · 0 invented entities

The model introduces no new physical entities; it relies on fitted circuit parameters, standard weak-inversion transistor equations, and a noise-process assumption. The main burden is the OU assumption for the stochastic event generator and the many parameters fitted from measurements.

free parameters (4)
  • Photoreceptor and buffer capacitances Cpd, Cfb, Cpr, Csf
    Estimated from SPICE and fine-tuned to match pixel noise PSD measurements (Section III).
  • Bias currents Ipd, Ipr, Isf
    Set from bias settings; Ipr is a bias current and Ipd is photocurrent. Used to compute gm and gs under weak inversion (Section II).
  • Transistor parameters kappa_fb, kappa_amp_n, kappa_sf, VA_amp_n, VA_amp_p
    Fitted or estimated from SPICE and measurements; used in the gm and Rout expressions in Section II.
  • Flicker noise magnitude
    Added to match low-frequency PSD in bright conditions; acknowledged as an adjustment without independent calibration (Section III).
assumptions (5)
  • domain assumption Weak inversion operation with gm = kappa*I/UT and gs = I/UT
    Used in Section II to express all small-signal parameters in terms of bias currents; valid for subthreshold MOSFET operation assumed for DVS pixels.
  • domain assumption Shot noise at Vsf is the sum of independent contributions from Ipd, Ipr, and Isf
    Equation in Section II; standard for transistor shot noise but neglects correlations and other noise sources.
  • domain assumption Shot noise can be modeled as an Ornstein-Uhlenbeck process (first-order low-pass filtered white noise)
    Section V: stated to be valid when Ipr >> Isf and there is a clear dominant pole between tau_pd and tau_sf. This is load-bearing for the stochastic event generator.
  • domain assumption Large-signal response can be approximated by linearizing the system at each timestep
    Section IV: the bilinear transform and per-timestep linearization are used; a standard quasi-static approximation, but no error bound is given.
  • standard math Bilinear transform discretization preserves the continuous-time behavior for the chosen timestep
    Section IV: a standard numerical integration method, but accuracy for large signals is not quantified.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Towards a physically realistic computationally efficient DVS pixel model." pith.science (2026). https://pith.science/paper/VONSLQMH

@misc{pith2026250507386,
  author       = {Pith},
  title        = {Pith review of: Towards a physically realistic computationally efficient DVS pixel model},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/VONSLQMH}},
  note         = {Machine review of arXiv:2505.07386}
}
read the original abstract

Dynamic Vision Sensor (DVS) event camera models are important tools for predicting camera response, optimizing biases, and generating realistic simulated datasets. Existing DVS models have been useful, but have not demonstrated high realism for challenging HDR scenes combined with adequate computational efficiency for array-level scene simulation. This paper reports progress towards a physically realistic and computationally efficient DVS model based on large-signal differential equations derived from circuit analysis, with parameters fitted from pixel measurements and circuit simulation. These are combined with an efficient stochastic event generation mechanism based on first-passage-time theory, allowing accurate noise generation with timesteps greater than 1000x longer than previous methods

Figures

Figures reproduced from arXiv: 2505.07386 by the authors.

Figure 1
Figure 1. A: DVS pixel architecture. B: Diagram of the proposed Dynamic Vision Sensor (DVS) pixel model with input Ipd and output events e. A. Contributions [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. Photoreceptor (PR) and source-follower (SF) buffer [PITH_FULL_IMAGE:figures/full_fig_p002_2.png] view at source ↗
Figure 3
Figure 3. Measured pixel PSD (faded lines) and proposed model (solid lines) for different on-chip illuminances for ( [PITH_FULL_IMAGE:figures/full_fig_p003_3.png] view at source ↗
Figures from the paper (3 more)
Figure 4
Figure 4. Figure 4: Model response to a step in photocurrent [PITH_FULL_IMAGE:figures/full_fig_p003_4.png]
Figure 5
Figure 5. Figure 5: Model response to a 1 ms pulse from 10 fA to 1 pA in photocurrent. (A) shows simulated signal at Vpr, (B) at Vsf, and (C) at Vdiff. (D) shows the generated ON and OFF events. 3.0 3.5 4.0 4.5 5.0 5.5 6.0 Threshold (relative to σ) 10−8 10−7 10−6 10−5 10−4 10−3 10−2 10−1 …
Figure 6
Figure 6. Figure 6: Noise event simulation with and without optimized stochastic event generation. (A) shows the event rate variation with [PITH_FULL_IMAGE:figures/full_fig_p004_6.png]

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

20 extracted references · 9 canonical work pages

  1. [1]

    Lichtsteiner et al

    P. Lichtsteiner et al. , JSSC, 2008. DOI: 0.1109/JSSC.2007.914337

    arXiv 2008

  2. [2]

    Gallego et al

    G. Gallego et al. , en, IEEE T-PAMI, 2020. DOI: 10.1109/TPAMI.2020.3008413

    arXiv 2020

  3. [5]

    Niwa et al

    A. Niwa et al. , in ISSCC, 2023. DOI: 10.1109/ISSCC42615.2023.10067566

    arXiv 2023

  4. [6]

    Serrano-Gotarredona et al

    T. Serrano-Gotarredona et al. , JSSC, 2013. DOI: 10.1109/JSSC.2012.2230553

    arXiv 2013

  5. [7]

    D. P. Moeys et al. , TBioCAS, 2018. DOI: 10.1109/TBCAS.2017.2759783

    arXiv 2018

  6. [8]

    Graca et al

    R. Graca et al. , in ESSERC, 2024. DOI: 10.1109/ESSERC62670.2024.10719521

    arXiv 2024

  7. [9]

    Rebecq et al

    H. Rebecq et al. , in CoRL, 2018. DOI: 10.5167/uzh-175993

    work page doi:10.5167/uzh-175993 2018

  8. [10]

    Hu et al

    Y . Hu et al. , in CVPRW, 2021. DOI: 10.1109/CVPRW53098.2021.00144

    arXiv 2021

Show all 20 references
  1. [11]

    Lin et al

    S. Lin et al. , in ECCV, 2022. DOI: 10.1007/978-3-031-20071-7 34

    2022 doi

  2. [12]

    Zhang et al

    Z. Zhang et al. , in ICRA, 2024. DOI: 10.1109/ICRA57147.2024.10609864

    2024

  3. [13]

    Jiang et al

    X. Jiang et al. , 2024. arXiv: 2411.12250 [cs.CV]

    2024 arXiv

  4. [14]

    Suess et al

    A. Suess et al. , in IISW, 2023. DOI: 10.60928/o8kk-2mew

    2023 doi

  5. [15]

    Mou et al

    X. Mou et al. , Electronic Imaging, 2024. DOI: 10.2352/EI.2024.36.7.ISS-289

    2024 doi

  6. [16]

    Joubert et al

    D. Joubert et al. , Front. Neurosci., 2021. DOI: 10.3389/fnins.2021.702765

    2021

  7. [17]

    Graca et al

    R. Graca et al. , in IISW, 2023. DOI: 10.60928/dlpf-irjd

    2023 doi

  8. [18]

    Graca et al

    R. Graca et al. , in CVPRW, 2023. DOI: 10.1109/CVPRW59228.2023.00423

    2023

  9. [19]

    Graca, ETH Zurich, 2024

    R. Graca, ETH Zurich, 2024. DOI: 10.3929/ethz-b-000683623

    2024 doi

  10. [20]

    Bibbona et al

    E. Bibbona et al. , Metrologia, 2008. DOI: 10.1088/0026-1394/45/6/S17

    2008 doi

  11. [21]

    M. T. Giraudo et al. , Commun. Stat. Simul. Comput. , 1999. DOI: 10 . 1080 / 03610919908813596

    1999

  12. [22]

    Graca et al

    R. Graca et al. , in IISW, 2021. DOI: 10.60928/n1k7-etj1

    2021 doi

Pith tools

Reviewed August 15, 2026 · model on record in the stance chip above.