{"id":"51399c9f-c1fd-425d-bc8c-f5e3beedfa92","arxiv_id":"2511.18037","paper_version":2,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":8,"one_line_summary":"A unified, calibrated noise model for hybrid event-frame sensors is implemented in H-ESIM, a simulator that generates realistic RAW frames and events and improves downstream frame interpolation and deblurring on real hybrid-sensor data.","lead":"The paper builds a statistical noise model for hybrid cameras that capture both ordinary frames and event signals on one chip, and turns it into a simulator that adds realistic sensor noise to synthetic video. It also calibrates the model on two real hybrid sensors and shows that training vision networks on its simulated data improves their performance on real sensor data.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Eq. 12's linear shot-noise term σ_shot=β3 I_c conflicts with Sec. 3's stated σ_shot² ∝ V̂_t; the EVS calibration's brightness dependence is a flexible fit, not the claimed physics.","rationale":"The paper's contribution rests on a physically grounded EVS noise model, and Eq. 12 is the quantitative core of the EVS calibration. If its shot-noise scaling is inconsistent with Sec. 3, the simulator's event statistics are not derived from the model but from an arbitrary regression. This is more fundamental than the missing-control issue: even a perfect transfer experiment would not validate a simulator whose noise generator uses a nonphysical scaling. I am not accusing the authors of dishonesty; this may be a typo or an unstated effective-voltage convention, but it needs to be resolved. The proposed test is feasible with data already described in the paper (static brightness sweeps). If the slope check supports linear scaling, the concern is withdrawn and the current CONDITIONAL verdict stands. The reader correctly identified the EVS side as the weak link; this concern pinpoints a concrete internal inconsistency within that side rather than relying only on generalization premises.","tokens_in":16918,"tokens_out":12424,"duration_ms":125267,"concrete_test":"Use the static calibration data to measure event probability P(I) across a wide brightness range. With the calibrated affine mapping V=β1 I_c+β2, compute from Eq. 11 the quantity A(V)=σ_n²V²≈2(σ_shot²+σ_DCSN²−2ρσ_shotσ_DCSN). Subtract the dark-current constant estimated from dark frames, then fit log(A_dark_subtracted) versus log V. A slope near 1 confirms σ_shot²∝V and contradicts Eq. 12's σ_shot=β3 I_c; a slope near 2 supports the linear form. Independently, re-fit Eq. 12 with σ_shot=β3√I_c and compare held-out brightness-level event-rate predictions; if the sqrt model predicts held-out P as well as or better than the linear model, the current calibration is a fit artifact rather than a physical scaling.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim requires that calibrated EVS event statistics generalize from static calibration scenes to the conditions where the simulator is used. The weakest point is Eq. 12, the quantitative core of EVS calibration. In Sec. 3, photon shot noise is defined with σ_shot² ∝ V̂_t (Poisson scaling). Combined with Eq. 11 in a static scene, σ_n² = 2(σ_shot²+σ_DCSN²−2ρσ_shotσ_DCSN)/(V̂+V_PD)². With physical shot-noise scaling, this event-noise variance decreases with brightness, so P=Q(θ/σ_n) should decrease with illumination. Yet Fig. 5(d) shows event probability increasing with brightness. The calibration resolves this by replacing the stated physics with σ_shot=β3 I_c in Eq. 12, making σ_shot roughly linear in the affine voltage (β1 I_c+β2). That is a regression form chosen to match the observed P(I), not a consequence of the noise model; it changes the predicted brightness scaling and can make σ_n increase with I. If the true shot-noise exponent is 1/2 rather than 1, the calibrated event simulator will not extrapolate to brightness levels, temperatures, or dynamics outside the calibration range, directly undermining 'statistically grounded.' The downstream VFI/deblurring improvements do not test this, since they only compare with/without H-ESIM fine-tuning and lack a control simulator.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes the first unified statistical noise model for hybrid event-frame sensors, jointly describing APS and EVS noise (shot, dark-current, fixed-pattern, quantization). It introduces a calibration pipeline that estimates noise parameters from real hybrid sensors (AlpsenTek GEN2 and Eiger), and presents H-ESIM, a simulator that generates synthetic RAW frames and events with calibrated noise statistics. The simulator is evaluated by fine-tuning video frame interpolation and deblurring networks on synthetic data and testing on real hybrid-sensor data, reporting improvements in no-reference quality metrics. The central claim is that calibrated, jointly modeled noise statistics enable realistic simulation that transfers to downstream tasks.","tokens_in":17404,"tokens_out":5128,"duration_ms":47119,"significance":"If the claims hold, the paper would provide a principled, reproducible simulation tool for the emerging hybrid event-frame sensor domain, potentially reducing the need for costly real-data collection for training. The release of an open NumPy/PyTorch simulator and the use of a 3200 fps input video dataset to avoid interpolation artifacts are concrete, useful contributions. However, the validation currently has important gaps: the EVS noise model is fit and tested on the same data, the transfer experiments lack a control simulator, and a stated physical scaling in Section 3 is contradicted by the calibration equation in Section 4. These issues bear directly on the 'statistically grounded' claim, so the significance is contingent on resolving them.","major_comments":[{"comment":"The shot-noise model is internally inconsistent. Section 3 states σ_shot² ∝ V̂_t (Poisson scaling), which implies σ_shot ∝ sqrt(β1 I_c + β2). Equation (12), however, sets σ_shot = β3 I_c, a linear dependence on intensity. This is not a minor notational slip: it changes the predicted brightness dependence of event probability. With the physical sqrt scaling and the static-scene Q-function, P = Q(θ/σ_n) should decrease as brightness increases (since σ_n ∝ 1/sqrt(V)), yet Figure 5(d) shows P increasing with brightness. The calibration thus substitutes a flexible regression form for the stated physics. The paper should either revise Section 3 to present the model actually used (empirical intensity-dependent shot noise) or extend the physical derivation to explain the observed increase. The current presentation undermines the 'statistically grounded' claim.","section":"§3 vs. §4, Eq. (12)"},{"comment":"The EVS noise 'validation' is in-sample. The parameters β_e are estimated by fitting Eq. (12) to observed event probabilities P(I_c) from static scenes, and Fig. 5 then compares model-generated probabilities to those same observed probabilities. This is a check of the fit, not an independent prediction. The same holds for APS in Fig. 4(d)–(g), where the fitted variance polynomial is evaluated on the calibration data. To support generalization, the authors should hold out brightness levels, scenes, or sensors and report prediction error, or validate the calibrated event statistics on dynamic scenes where the signal part S is nonzero.","section":"§4, Event Noise Calibration; §6.1, Fig. 5"},{"comment":"The downstream transfer experiments do not isolate the contribution of the calibrated noise statistics. Fine-tuning with H-ESIM is compared only against not fine-tuning; there is no control fine-tuned with an existing simulator (e.g., ESIM or v2e) or with an uncorrupted/ideal event stream. Consequently, the observed improvements could stem from the 3200 fps input, the specific event-generation pipeline, or joint RAW+event simulation, rather than from the calibrated noise parameters. Additionally, Table 1 reports BRISQUE for Eiger HR-INR(w) as 45.43 vs. 31.45 for (w/o), yet the text claims 'lower distortion' — a direct contradiction that must be addressed.","section":"§6.2, Tables 1–2"},{"comment":"The parameter set β_e = {β0,...,β5} in Eq. (12) is not identifiable from P(I) alone. The right-hand side is invariant under rescaling transformations (e.g., multiplying β0 and β1 by reciprocal constants, or scaling β3 and β4 together and adjusting β0). The paper does not discuss this gauge freedom or impose constraints. This matters because the calibrated β's are presented as interpretable physical noise parameters (shot coefficient, dark-current coefficient, correlation). At minimum, the authors should state which parameters are identifiable and how the remaining degrees of freedom are fixed, or drop the interpretability claim for those parameters.","section":"§4, Eq. (12), parameter identifiability"}],"minor_comments":[{"comment":"Parameter names are inconsistent: the text says 'σ_shot ≈ β4 I_c, σ_DCSN ≈ β5, and correlation ρ≈β6', but Section 4 defines σ_shot = β3 I_c, σ_DCSN = β4, ρ = −β5. Please correct to avoid confusion.","section":"§5, EVS Simulator, step (3)"},{"comment":"The mapping from I_c to V̂ is affine (V̂+V_PD ≈ β1 I_c + β2). Since σ_shot in Eq. (12) is linear in I_c, the denominator in Eq. (12) can be re-expressed directly in terms of V̂. Clarify whether β3 is a voltage-domain or intensity-domain coefficient, and ensure all references (including Sec. 5) use the same convention.","section":"§4, Eq. (11)–(12)"},{"comment":"The improvement for MAER is marginal (CLIP-IQA 0.3297→0.3370, MUSIQ 18.88→19.05, NRQM 5.038→5.064). Reporting statistical significance or confidence intervals would help assess whether the fine-tuning benefit is real for that model.","section":"§6.2, Table 2"}],"recommendation":"major_revision","confidential_remarks":"The paper addresses a timely topic and the artifact (simulator, calibration pipeline, high-speed dataset) is potentially valuable. The main concern is that the central 'statistically grounded' claim is not yet supported: the EVS model has an internal inconsistency between the stated physics and the calibrated regression form, and the experimental evaluation does not isolate the noise-model contribution. These are fixable with revised presentation and additional control experiments, so I recommend major revision rather than rejection. If the authors can add a control simulator and reconcile the shot-noise scaling, the paper could become a solid contribution to the event-camera community."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Hi [Colleague],\n\nTwo things to know up front. First, this is the paper to read if you care about hybrid event-frame sensors: it actually models APS and EVS jointly, calibrates on two real sensors, and ships a simulator with an open ISP. Second, the EVS noise model has an internal contradiction that the authors don't acknowledge: Section 3 says shot noise scales as σ_shot² ∝ V̂_t (Poisson), but Eq. 12 uses σ_shot = β3 I_c, a linear-in-intensity regression. Those are different scalings. The calibration is therefore a flexible fit, not the physics they claim, and it's why Fig. 5(d) can show event probability increasing with brightness while the stated Poisson scaling would predict the opposite. This is a genuine soft spot, not a nitpick.\n\nWhat's actually new and good: the unified treatment of APS and EVS through the shared irradiance I_c, the per-CFA-position APS calibration (which captures row noise and color-dependent variance), and the H-ESIM simulator with a transparent NumPy ISP. The APS calibration on GEN2 and Eiger looks careful—the variance fits and the fixed-pattern separation are plausible. The paper also collects and releases (apparently) calibration data and code, which is valuable.\n\nThe weaknesses, in proportion: (1) The EVS validation is in-sample by construction. Fig. 5 compares the fitted model to the same event probabilities used to estimate β_e. That's a check of the fit, not a prediction. (2) The downstream VFI/deblurring experiments lack a control: fine-tuning on ESIM or v2e, or on real data. Without that, 'strong transfer' is unproven. (3) There are notation slips in the simulator section (σ_shot≈β4, σ_DCSN≈β5, ρ≈β6 vs Eq. 12's β3/β4/β5), and the affine voltage mapping is asserted without justification. These are fixable but exactly the kind of thing that slows down reproducibility.\n\nOn balance, I'd send this to peer review. The integration and resources are worthwhile, and the EVS consistency issue is exactly what a sharp referee should catch. But it needs major revision: reconcile the shot-noise scaling, add an independent EVS validation (e.g., held-out brightness levels or temperatures), and run a proper control in the transfer experiments.\n\nBest,\n[Your name]","headline":"Solid APS side, shaky EVS core: the unified-noise idea is right, but Eq. 12 trades away the stated physics and the transfer claims lack a control.","tokens_in":17865,"tokens_out":4872,"would_cite":true,"duration_ms":40922,"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":"The paper claims that a single statistical noise model can describe both the RAW frames and the event stream of a hybrid event-frame sensor, and that a simulator calibrated from that model yields synthetic data that improves real-sensor vid","keywords":["hybrid event-frame sensors","APS EVS unified noise model","event camera noise calibration","event simulation","RAW frame synthesis","Q-function event probability","video frame interpolation","video deblurring"],"falsifier":"Record a static grayscale ramp at several exposure times and temperatures, count per-pixel event probabilities, and fit Q^{-1}(P) against the affine-plus-square-root form of Eq. 12. A systematic nonlinearity in I_c, or a change in event rate with temperature at fixed brightness, would falsify the constant-dark-current and fixed-correlation assumptions.","tokens_in":16781,"feed_emoji":"📷","tokens_out":5351,"duration_ms":44348,"temperature":0.7,"pith_summary":"The paper sets out to establish that hybrid event-frame sensors—chips that output both ordinary intensity frames and asynchronous events from one pixel array—can be described by a single statistical noise model rather than two separate ones. It derives a formulation in which photon shot noise, dark current, fixed-pattern noise, and quantization noise appear in both the APS and EVS paths, with the event triggering probability linked to illumination through the Q-function. From this model it builds a calibration pipeline that estimates noise parameters from static dark and brightness patterns, and then a simulator, H-ESIM, that generates RAW frames and events under those calibrated statistics. If the model is right, synthetic data from H-ESIM carries the sensor's real noise behavior, and the paper shows that fine-tuning interpolation and deblurring networks on such data improves their results on real hybrid-sensor recordings.","feed_headline":"One noise model unites frames and events in hybrid sensors","feed_subtitle":"Calibrated statistics let simulated RAW frames and events improve real deblurring and interpolation.","key_machinery":"The mechanism that carries the argument is the Q-function identity σ_n^2 = (θ / Q^{-1}(P))^2, which turns the discrete ON/OFF output of an event pixel into a measurable noise variance. Around this identity the model uses a shared ideal signal I_c, a second-order polynomial Var(N_a) = β0 + β1 I_c + β2 Δt + β3 I_c^2 + β4 I_c Δt + β5 Δt^2 for APS noise variance, and an affine voltage mapping V̂ + V_PD ≈ β1 I_c + β2 to connect the APS intensity domain to the EVS logarithmic voltage domain. Together these allow a calibration pipeline to convert observed event probabilities and frame variances into a parameter set that drives the simulator's noise injection and threshold comparison.","core_discovery":"The core claim is that both pixel types in a hybrid sensor—the integrating APS pixel and the differential, thresholding EVS pixel—obey the same noise-generating physics, so both can be written as Gaussian noise terms added to a shared ideal signal. The key relationship is P+ = P− = Q(θ/σ_n) for a static scene, which means the probability of a noise-triggered event directly reveals the event noise variance. Calibrating static multi-brightness frames therefore yields the shot-noise, dark-current, and fixed-pattern parameters for both modalities, and the simulator injects those same statistics when generating frames and events. The authors validate this on two hybrid sensors by matching measure","pith_inferences":["Because the calibrated model ties event rates to absolute brightness, the same pipeline could be inverted to estimate scene irradiance from event counts alone, which the paper does not claim to do.","The per-position Quad-Bayer coefficients expose spatially structured noise at the color-filter level, suggesting a demosaicing or denoising network could be designed to consume those calibrated positions as side information.","The Gaussian model for event noise will likely need a Poisson or Gamma replacement at very low photon counts; testing H-ESIM in that regime would directly probe the model's range of validity."],"forward_implications":["Calibrated noise parameters from a real hybrid sensor can be reused to synthesize RAW frames and events for that sensor without hand-tuned thresholds or heuristic event rules.","Fine-tuning video frame interpolation and deblurring networks on H-ESIM synthetic data improves their perceptual quality on real hybrid-sensor sequences.","The unified model separates illumination-dependent, exposure-dependent, and fixed noise, so it predicts how sensor noise changes with brightness and exposure time.","The calibration has to be performed per sensor and per layout; the two sensors studied show different row noise and color-filter-dependent variance, so transfer across sensor models is not automatic."],"fun_headline_variants":["Hybrid sensor noise: one model, two pixel types","Unified noise physics for frames and events in hybrid sensors","Calibrated noise model improves hybrid sensor simulation","One noise model for hybrid sensors: from calibration to simulation"],"cache_read_input_tokens":2304,"weakest_assumption_plain":"The calibration assumes that in static scenes the event signal S is zero and the mean event noise μ_n is zero, and that the mapping from APS intensity to EVS voltage is affine; if these fail under motion, low light, or temperature change, the calibrated event statistics and the simulator's events will no longer match the real sensor.","fun_headline_variants_meta":{"raw":{"variants":["Hybrid sensor noise: one model, two pixel types","Unified noise physics for frames and events in hybrid sensors","Calibrated noise model improves hybrid sensor simulation","One noise model for hybrid sensors: from calibration to simulation"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000282,"raw_usage":{"total_tokens":1497,"prompt_tokens":729,"completion_tokens":768,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":473,"completion_tokens_details":{"reasoning_tokens":703}},"tokens_in":473,"tokens_out":768,"duration_ms":6839,"temperature":1.0,"reasoning_tokens":703,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-03T20:48:43.198336+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Record a static grayscale ramp at several exposure times and temperatures, count per-pixel event probabilities, and fit Q^{-1}(P) against the affine-plus-square-root form of Eq. 12. A systematic nonlinearity in I_c, or a change in event rate with temperature at fixed brightness, would falsify the constant-dark-current and fixed-correlation assumptions.","supporting_citations":[],"review_version":1}