{"id":"d13fb45e-8d28-4b23-be42-e50b6df8a3d1","arxiv_id":"2411.12250","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":2,"one_line_summary":"A video-to-events simulator that over-samples frames and applies a brightness-dependent low-pass filter produces synthetic events that train segmentation and reconstruction networks closer to real event data than prior simulators.","lead":"ADV2E is a new simulator that converts ordinary video frames into the event-camera data used to train vision algorithms, adding a brightness-dependent analog filter to match real pixel circuitry. If it holds up, it could give event-camera researchers a cheap way to generate realistic training data from ordinary videos.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Eq. (4) uses α as the input coefficient of the discretized low-pass filter; the exact first-order discretization with unity DC gain requires 1−e^{−α}. At the stated settings (24 FPS, K=L=10, 250 Hz cutoff) α≈0.65, giving a ~37% DC-gain error that worsens for brighter pixels.","rationale":"The reader's weakest assumption points to the unverified first-order, brightness-proportional filter model and the missing supplementary derivation. My reading agrees that this is a real soft spot, but I found a more concrete and independently checkable problem: even granting the first-order model, Eq. (4) is not the correct discretization of that model. The input coefficient should be 1−e^{−α}, not α, for the filter to preserve the stated DC gain. At the paper's own operating point the discrepancy is large, and the brightness-proportional cutoff makes the error scene-dependent, exactly in the high-contrast regimes where the paper claims improvement. I credit the consistent empirical gains across the three tasks, but they do not by themselves validate the mechanism if the mechanism as implemented is miscalibrated. The proposed check—re-running with the corrected coefficient—would settle whether the reported advantages survive. I therefore keep the reader's CONDITIONAL verdict: the paper is promising and the concern is testable, but the current text does not establish the analogue-fidelity claim as written.","tokens_in":11220,"tokens_out":5872,"duration_ms":69980,"concrete_test":"Re-run Tables 1–3 with the only change Y ← e^{−α}·Y + (1−e^{−α})·I′, keeping K=10, L=10, and all interpolation, noise, and training settings fixed, and adopting one consistent Hz-vs-rad/s convention for ω₀. If the temporal-bin distance in Table 1 ceases to favour ADV2E, or segmentation mIoU and reconstruction MSE/SSIM no longer favour ADV2E, the headline gains are attributable to the gain error in Eq. (4) rather than to analogue-circuit modelling. As a unit-level sanity check, feed a step input and verify that the steady-state filtered output matches the unity DC gain of the intended first-order low-pass filter.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The paper's central mechanism is Eqs. (2)–(4): a first-order low-pass filter with brightness-dependent cutoff, discretized as Y = e^{−α}·Y + α·I′, where α = ω₀·T_b/(KL). For the continuous system H(s) = ω₀/(s+ω₀), the exact zero-order-hold update with unity DC gain is Y ← e^{−α}·Y + (1−e^{−α})·I′. Eq. (4) uses α instead of 1−e^{−α}; these agree only when α ≪ 1. With the paper's own settings (§4.1: f_b = 24 FPS, K = 10, L = 10), Δt = 1/2400 s; interpreting the stated 250 Hz cutoff as ω₀ = 2π·250 rad/s gives α ≈ 0.65, and the filter's DC gain is α/(1−e^{−α}) ≈ 1.37 instead of 1. If ω₀ is instead read as 250 rad/s, the paper is ambiguous about Hz versus rad/s and the error is smaller but still systematic. Because ω₀ ∝ I, high-brightness pixels in the high-contrast validation scenes can have α > 1, making the DC gain approach α and producing a large brightness-dependent amplitude distortion. Thus the claimed 'faithful' analogue filtering, and the reported gains over v2e, may be artifacts of the discretization in Eq. (4) rather than evidence for the true DVS pixel response. The absent supplementary derivation cannot repair Eq. (4) as printed.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes ADV2E, a video-to-events simulator that incorporates the analogue low-pass filtering behavior of DVS pixel circuits into the event generation pipeline. The method has two main components: continuity sampling, which over-samples interpolated video frames by a factor K to increase the effective sampling rate, and an analogue low-pass filter whose cutoff frequency is assumed proportional to instantaneous brightness, discretized with an impulse-invariance-style update. The authors evaluate ADV2E against four baselines (vid2e, v2e, v2ce, DVS-Voltmeter) on three tasks: direct event comparison using temporal-bin distance on the DAVIS240C 'urban' sequence, semantic segmentation on DDD17 using Ev-Segnet, and image reconstruction on DAVIS240C using E2VID trained on GOPRO-simulated events. ADV2E reports the best temporal-bin distance (49.58), best mIoU (51.20), best MSE (0.02), and best SSIM (0.55), with mixed LPIPS results. The paper claims that these results validate that modeling analogue circuit properties improves the realism of simulated events and sim-to-real generalization.","tokens_in":11536,"tokens_out":3945,"duration_ms":41753,"significance":"If the central modeling claim is correct, ADV2E addresses a genuine gap in video-to-events simulation: existing simulators treat event generation as a purely logical thresholding process and largely ignore the analogue low-pass filtering that is physically present in DVS pixels. The experimental protocol is appropriate and externally grounded: training on simulated events and testing on real event datasets (DDD17, DAVIS240C) is a meaningful sim-to-real test, and the temporal-bin distance on a real DAVIS sequence provides a direct event-level comparison. The paper also honestly notes in §2.1 that the exact filter order of the DVS pixel is unclear, and it names the assumptions underlying the first-order model. However, the current manuscript does not yet support the strong causal claim that the observed improvements come from the specific analogue filtering mechanism: the discretization in Eq. (4) is not the exact first-order ZOH update, the two proposed components are never ablated separately, and no error bars or statistical significance tests are reported. The contribution is promising but needs substantial strengthening before the claims can be accepted.","major_comments":[{"comment":"The discretization of the first-order low-pass filter is not the exact zero-order-hold update for H(s)=ω0/(s+ω0). With sample interval Δt=T_b/(KL), the correct recurrence with unity DC gain is Y ← e^{-α}Y + (1−e^{-α})I', where α=ω0Δt, whereas Eq. (4) uses Y = e^{-α}Y + αI'. These agree only when α≪1. At the paper's own settings (§4.1: 24 FPS, K=10, L=10), if the stated 250 Hz cutoff is interpreted as ω0=2π·250 rad/s, then α≈0.65 and the DC gain is α/(1−e^{-α})≈1.37, a 37% amplitude error that grows for brighter pixels because ω0∝I. If ω0 is instead read as 250 rad/s, the paper is ambiguous about Hz versus rad/s and the error is smaller but still systematic. This directly undermines the claim that the filter 'faithfully' replicates the analogue DVS response, and it means the reported gains over baselines may be contaminated by a brightness-dependent amplitude distortion. The derivation for the first-order model is also deferred to a supplementary file that is absent from this preprint, so the central model cannot be fully checked as printed.","section":"§3.2.2, Eq. (4)"},{"comment":"The two main contributions—continuity sampling (K) and the brightness-dependent analogue low-pass filter—are never ablated separately. K=10 is set in §4.1, and all other parameters are inherited from v2e/vid2e, but no experiment removes the time-varying cutoff or varies K to show that the improvements in Table 1, Table 2, and Table 3 are due to the analogue filtering mechanism rather than simply to the higher effective sampling rate. Without such an ablation, the causal claim that 'the gap between analogue pixel circuit and discrete video frames' is the source of improvement is not established.","section":"§4.1 and Tables 1–3"},{"comment":"No error bars, standard deviations, or numbers of random seeds are reported for any of the three evaluations. The margins in the headline results are small—mIoU 51.20 versus 50.60, SSIM 0.55 versus 0.53—and in Table 3 the LPIPS result is not even the best (0.44 versus vid2e's 0.41, and per-sequence LPIPS is worse for shapes6dof, 0.48 versus DVS-Voltmeter's 0.28). These differences may be within run-to-run variability when training Ev-Segnet and E2VID from scratch, so the current evidence does not support the strength of the claims without a significance analysis.","section":"Tables 1–3 (all experiments)"},{"comment":"The proportionality ω0(t) ∝ I(t) is a load-bearing assumption, but the proportionality constant is never specified, calibrated, or studied for sensitivity. The paper states in §4.1 that K is the only parameter requiring adjustment, yet the brightness-to-cutoff mapping introduces at least one additional free parameter whose value is not reported. Since this parameter directly controls the magnitude of α in Eq. (4) and hence the amplitude distortion noted above, the quantitative results are not reproducible without this information.","section":"§3.2.2, §4.1"}],"minor_comments":[{"comment":"The phrase 'as an infantile sensor' is nonstandard and should be replaced with, for example, 'as a nascent sensor technology'.","section":"§1, Abstract"},{"comment":"The Nyquist argument is internally inconsistent: for a 250 Hz cutoff the text says the minimum sampling rate is 500 FPS, which with 24 FPS input requires an interpolation factor of at least 20.9, not 'at least 16x' as stated.","section":"§3.2.1"},{"comment":"The names 'DA VIS240C' and 'DVS-V oltmeter' appear with broken spacing throughout, presumably from LaTeX control sequences; these should be rendered as 'DAVIS240C' and 'DVS-Voltmeter'.","section":"§4.1, captions and table headers"},{"comment":"Figure 3 is never cited in the body text; either add a reference at the appropriate point in §3.2.2 or remove the figure.","section":"Figure 3"},{"comment":"The text says ADV2E achieves the second-best LPIPS, which matches Table 3, but the qualitative claim that ADV2E produces the most realistic events should be tempered given that vid2e has better average LPIPS and several per-sequence LPIPS comparisons favor other methods.","section":"§4.4 and Table 3"},{"comment":"Reference [17] appears to be mislabeled: it cites a paper by Kim et al. but with the journal and page numbers of the DVS paper [21]; please verify the citation details.","section":"References"}],"recommendation":"major_revision","confidential_remarks":"The paper is within the scope of a computer vision venue and the sim-to-real experimental setup is a good choice. My main concern is not the experimental idea but the correctness and identifiability of the core filter model: the discretization error in Eq. (4), the absence of ablations separating K from the filter, and the lack of uncertainty quantification all need to be addressed before the claims can be supported. If the authors can fix the discretization, add ablations, and report variance, a revised version could be publishable."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Short version: ADV2E is a reasonable step forward for video-to-events simulation — over-sampling plus a brightness-dependent low-pass filter is a sensible extension of v2e, and the downstream numbers (segmentation mIoU 51.20, reconstruction MSE 0.02, SSIM 0.55) are consistently better than the four baselines. But the paper's central 'faithful analogue' claim has a load-bearing math issue, and the experimental section is missing ablations, error bars, and code. I'd send it to a serious referee, but the referee should push hard on Eq. (4) before accepting the mechanism.\n\nThe new thing here is the combination: continuity sampling (K=10) on top of v2e's frame interpolation, with the low-pass cutoff updated at every over-sampled step. That is a genuine change, and the evaluation on two real-world downstream tasks is a good-faith attempt to show sim-to-real value. The authors are also honest that the exact filter order is unclear (§2.1), and they cite the relevant v2e and DVS biasing literature.\n\nWhere it gets shaky: Eq. (4) uses Y ← e^{−α}Y + αI'. For the continuous filter H(s)=ω0/(s+ω0), a unity-DC-gain discretization (ZOH or impulse invariance with gain correction) requires the input coefficient to be 1−e^{−α}. At their own settings (24 FPS, K=L=10, 250 Hz cutoff), α≈0.65, so the DC gain of the printed filter is about 1.37 — and it grows for brighter pixels. That is a brightness-dependent amplitude distortion, not a faithful analogue model. If they want to keep impulse invariance, they should normalize; if they want to defend the gain, they need to say so. The stress-test claim that the reported gains could be partly an artifact of this distortion is not crazy. I checked the arithmetic; the concern lands.\n\nThe other soft spots are what you'd expect: no ablation separating K from the brightness-dependent filter, no error bars on the segmentation/reconstruction numbers, a single video for direct comparison, and the supplementary derivation for the filter order is absent. None of these are fatal on their own, but together with Eq. (4) they make the central mechanism under-supported.\n\nBottom line: the paper deserves peer review — the idea is useful and the experiments are consistent — but it needs a revision that fixes or justifies Eq. (4), adds ablations, releases code/data, and supplies the missing filter derivation. I wouldn't cite it in current form.","headline":"A plausible simulator extension with consistent downstream gains, but Eq. (4) has a real DC-gain error that undermines the 'faithful analogue' claim; worth a serious referee, not citable as-is.","tokens_in":12107,"tokens_out":3890,"would_cite":false,"duration_ms":38916,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"This paper claims that modelling DVS pixels as brightness-dependent low-pass filters makes synthetic events realistic enough that networks trained on them transfer to real event cameras.","keywords":["event cameras","Dynamic Vision Sensor","event simulation","video-to-events","low-pass filter","continuity sampling","sim-to-real generalization","semantic segmentation"],"falsifier":"Point a real event camera at a screen that suddenly jumps from dark to bright, and record the timestamps of the resulting events. If the event rate after the jump is not a single exponential whose time constant shrinks as the brightness rises, the first-order brightness-proportional filter model is wrong.","tokens_in":10959,"feed_emoji":"⚡","tokens_out":8506,"duration_ms":80333,"temperature":0.7,"pith_summary":"Event cameras do not record frames; each pixel continuously emits a timestamped event when logarithmic brightness crosses a threshold. Simulators that turn ordinary video into synthetic events usually model only this logical rule, and their events drift from real camera output, especially in high-contrast scenes. This paper argues that the missing ingredient is the analogue pixel circuit, which behaves like a first-order low-pass filter whose cutoff frequency rises and falls with instantaneous brightness. The authors build ADV2E, which over-samples brightness changes and filters them through that time-varying model, and show that networks trained on its events transfer to real event data in semantic segmentation and image reconstruction. If true, existing frame-based video datasets become usable for training event-camera networks without costly real-event collection.","feed_headline":"Analogue-aware simulator narrows sim-to-real gap for event cameras","feed_subtitle":"Brightness-tracking analogue filtering and over-sampling make synthetic events closer to real DVS output.","key_machinery":"The load-bearing object is a time-varying first-order low-pass filter used as a model of the DVS pixel's analogue front-end. Its cutoff frequency $\\omega_0(t)$ is taken to be proportional to instantaneous brightness $I(t)$, so bright regions filter faster than dark ones; the filter is discretised by impulse invariance, giving the recurrence $Y=e^{-\\alpha}Y+\\alpha I'_{n,l,k}$. The companion mechanism is continuity sampling: each interpolated frame is linearly over-sampled $K$ times so that the effective sampling rate $KLf_b$ exceeds the Nyquist rate implied by the filter's cutoff, which prevents aliasing and lets the cutoff frequency track brightness within each short interval.","core_discovery":"The paper's central claim is that the visually obvious difference between synthetic and real events, especially at light-dark edges, is caused less by frame-rate interpolation than by the analogue low-pass filtering inside the DVS pixel. In a real DVS, brightness changes pass through a cascode feedback loop and a source follower whose finite bandwidth delays the signal; the paper models this delay as a first-order low-pass filter with system function $H_a(s)=\\omega_0(t)/(s+\\omega_0(t))$, where the cutoff frequency $\\omega_0(t)$ is proportional to the current brightness. To apply this model to discrete frames without aliasing, ADV2E first linearly over-samples every interpolated frame by a factor $K$, making the effective sampling rate $KLf_b$, and then runs the filter through the recurrence $Y=e^{-\\alpha}Y+\\alpha I'_{n,l,k}$ with $\\alpha=\\omega_0^{n,l,k}T_b/(KL)$. The paper reports that this procedure yields synthetic events with the smallest temporal-bin distance to real DVS recordings (49.58) among the compared simulators, and that segmentation and reconstruction networks trained on ADV2E events achieve the best mIoU (51.20) and best MSE (0.02) and SSIM (0.55) on real test data.","pith_inferences":["Beyond the paper, the same analogue-first approach suggests that other DVS non-idealities, such as leakage noise, threshold mismatch, and temperature-dependent bandwidth, could be simulated by the same continuity-sampling discretisation rather than bolted on as post-processing.","A testable extension the paper does not run is a controlled step-response experiment: exposing a real DVS to a sudden uniform brightness step and measuring the event-rate decay should show an exponential tail with time constant inversely proportional to brightness; a different functional form would locate where the first-order assumption breaks.","An ablation that varies the over-sampling factor $K$ at fixed interpolation $L$ would isolate how much of the reported improvement comes from alias removal versus the brightness-dependent cutoff itself."],"forward_implications":["Synthetic events from ordinary APS video become reliable training data for event-camera networks, reducing dependence on scarce and expensive real event datasets.","High-contrast scenes, where prior simulators show events abruptly vanishing, should retain realistic event tails caused by the brightness-dependent filter delay.","The simulator's only tunable parameter is the continuity sampling factor $K$, so the same pipeline can be applied to any frame-based video without per-scene calibration.","If the first-order model is right, networks trained on ADV2E events should keep their advantage as more high-contrast driving and indoor sequences are added."],"supporting_citations":[{"why":"It supplies the empirical anchor that nominal DVS latency is about 4 ms, implying a cutoff frequency near 250 Hz and a brightness-proportional cutoff.","marker":"[6]"},{"why":"It is the baseline simulator whose interpolation, log transform, and noise pipeline the proposed simulator inherits and whose low-pass filtering it refines.","marker":"[14]"},{"why":"It is the first of three cited sources for approximating the pixel's analogue behaviour as a low-pass filter.","marker":"[11]"},{"why":"It is the second cited source for the low-pass approximation of the pixel's analogue behaviour.","marker":"[12]"},{"why":"It is the third cited source for the low-pass approximation of the pixel's analogue behaviour.","marker":"[24]"},{"why":"It is the original pixel description that grounds the analysis of the cascode feedback loop and source follower.","marker":"[21]"},{"why":"It is the sampling theorem that fixes the required minimum sampling rate and motivates continuity sampling.","marker":"[33]"},{"why":"It is an adaptive-sampling simulator used as a comparison baseline for timestamp accuracy.","marker":"[31]"}],"fun_headline_variants":["ADV2E: Analog pixel filtering makes synthetic events realistic","Event simulator models pixel's analog low-pass filter","ADV2E: Discrete frames meet analog circuit in event sim","Analog delay in DVS pixels explains sim-to-real gap","Over-sampled frames plus analog filter make real-like events"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The argument rests on treating the DVS pixel's analogue behaviour as a first-order low-pass filter whose cutoff frequency is directly proportional to instantaneous brightness; the paper notes the exact filter order is unknown and defers the derivation to a supplementary file absent from this preprint.","fun_headline_variants_meta":{"raw":{"variants":["ADV2E: Analog pixel filtering makes synthetic events realistic","Event simulator models pixel's analog low-pass filter","ADV2E: Discrete frames meet analog circuit in event sim","Analog delay in DVS pixels explains sim-to-real gap","Over-sampled frames plus analog filter make real-like events"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.00064,"raw_usage":{"total_tokens":2987,"prompt_tokens":1025,"completion_tokens":1962,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":641,"completion_tokens_details":{"reasoning_tokens":1879}},"tokens_in":641,"tokens_out":1962,"duration_ms":16631,"temperature":1.0,"reasoning_tokens":1879,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-12T17:46:04.135819+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Point a real event camera at a screen that suddenly jumps from dark to bright, and record the timestamps of the resulting events. If the event rate after the jump is not a single exponential whose time constant shrinks as the brightness rises, the first-order brightness-proportional filter model is wrong.","supporting_citations":[{"cited_title":"V2e: From video frames to realistic dvs event camera streams","cited_arxiv_id":null,"evidence_quote":"It supplies the empirical anchor that nominal DVS latency is about 4 ms, implying a cutoff frequency near 250 Hz and a brightness-proportional cutoff."},{"cited_title":"v2e: From video frames to realistic dvs events","cited_arxiv_id":null,"evidence_quote":"It is the baseline simulator whose interpolation, log transform, and noise pipeline the proposed simulator inherits and whose low-pass filtering it refines."},{"cited_title":"Optimal biasing and physical limits of dvs event noise","cited_arxiv_id":null,"evidence_quote":"It is the first of three cited sources for approximating the pixel's analogue behaviour as a low-pass filter."},{"cited_title":"Shining light on the dvs pixel: A tutorial and discussion about biasing and optimization","cited_arxiv_id":null,"evidence_quote":"It is the second cited source for the low-pass approximation of the pixel's analogue behaviour."},{"cited_title":"Demystifying event-based sensor biasing to optimize signal to noise for space domain awareness","cited_arxiv_id":null,"evidence_quote":"It is the third cited source for the low-pass approximation of the pixel's analogue behaviour."},{"cited_title":"A 128× 128 120 db 15 µs latency asynchronous temporal con- trast vision sensor","cited_arxiv_id":null,"evidence_quote":"It is the original pixel description that grounds the analysis of the cascode feedback loop and source follower."},{"cited_title":"Communication in the presence of noise","cited_arxiv_id":null,"evidence_quote":"It is the sampling theorem that fixes the required minimum sampling rate and motivates continuity sampling."},{"cited_title":"Esim: an open event camera simulator","cited_arxiv_id":null,"evidence_quote":"It is an adaptive-sampling simulator used as a comparison baseline for timestamp accuracy."}],"review_version":1}