{"id":"f9070d91-4931-46d7-9c24-fbf8e070eb5b","arxiv_id":"2607.25281","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":7.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":3,"one_line_summary":"An event-camera Shack-Hartmann wavefront sensor driven by modulated illumination recovers high-dynamic-range static wavefronts and tracks dynamic centroids at kilohertz rates on a CPU.","lead":"The paper builds a Shack-Hartmann wavefront sensor from an event camera and a flickering LED, so the camera's event timing carries spot brightness and position. This lets one sensor do both high-dynamic-range static metrology and kilohertz dynamic wavefront tracking, at CPU speed.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Dynamic accuracy is validated only against a 30 fps/20 ms frame reference, so the kHz tracking error claim may reflect agreement with an aliased/averaged ground truth rather than true high-frequency accuracy.","rationale":"The reader's weakest_assumption identifies exactly the same load-bearing concern: the dynamic quantitative evaluation uses a 30 fps frame-based centroid as ground truth for a system claimed to resolve sub-millisecond spot motion. This is not a minor reproducibility issue; it directly undermines the central claim of kilohertz-rate tracking accuracy. If the true high-frequency error is larger, the dynamic results reported in Tables 1 and Fig. 5e would not support the advertised capability, even though the static reconstruction and throughput results might remain valid. I do not see a more important weakness in the paper. The static comparison in Sec. 3A lacks error bars and the 'bright central' ground truth is not fully described, but those affect the strength of the static claim less than the dynamic validation gap affects the unified-framework claim. The paper does present concrete hardware implementation, a self-consistent throughput analysis, and a direct comparison against EBWFNet, which are independent supports. The dynamic validation issue is addressable with a high-speed reference or a known-trajectory experiment, so the appropriate disposition remains CONDITIONAL rather than REJECT. Since my read agrees with the reader's verdict, no change is needed.","tokens_in":13070,"tokens_out":3590,"duration_ms":38271,"concrete_test":"Use a calibrated piezo-driven tip/tilt mirror or fiber tip to move a focal spot on a known high-bandwidth trajectory (e.g., 500 Hz sinusoid, 0.5–2 px amplitude, plus a slow drift), and record simultaneously with EvTem-SHWFS (modulation on) and a 1 kHz frame camera (or use the commanded trajectory as ground truth). Compare EvTem centroids at native 500 Hz/1 kHz rate to the high-speed reference. If RMS error stays below ~0.3 px, the dynamic claim holds; if it is substantially larger, the reported 0.16–0.26 px errors are artifacts of comparing to a 30 Hz averaged GT. A synthetic event-stream simulation with known centroids would be a cheaper complementary check.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central dynamic claim — 500 Hz–1 kHz centroid tracking with 0.16–0.26 px errors — rests on a proxy ground truth that cannot see the high-frequency motion the sensor claims to resolve. In Sec. 3B, the reference is a 30 fps frame camera with 20 ms exposure (Fig. 4b); in Sec. 3C the same 30 Hz framing is used for turbulence (Fig. 5a). A 20 ms exposure averages spot motion over 10 modulation cycles at 500 Hz, and 30 Hz sampling aliases any content above 15 Hz. The reported errors are computed 'at GT timestamps' (Fig. 4d) and therefore measure agreement with the slowly varying, time-averaged reference, not instantaneous tracking error. Sub-millisecond oscillations are shown qualitatively in Fig. 4c but never quantitatively checked. The 0.16 px mean being close to the 0.14 px calibration floor is consistent with this reading: both sensors agree on the low-frequency centroid, while the claimed new capability (high-frequency fidelity) is untested. The same applies to the turbulence result: 0.26 ± 0.06 px is against a 30 Hz frame reference, so it does not establish 1 kHz accuracy.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper presents EvTem-SHWFS, a Shack–Hartmann wavefront sensor built on an event camera with active temporal modulation of illumination. In a static regime, low-frequency sine modulation encodes local focal-spot intensity into event latencies, permitting one-shot high-DR intensity reconstruction; in a dynamic regime, high-frequency square-wave modulation produces alternating-polarity event pairs that are filtered (PFF) and tracked in real time. The authors report a theoretical 260 dB DR at 20 Hz, static wavefront RMSE reductions of 59–70% versus a frame-based sensor, dynamic centroid errors of 0.18 px (alignment) and 0.26 px (turbulence), and a per-sub-aperture CPU throughput of 420,737 Hz. The comparison includes a public and a fine-tuned passive event-based baseline (EBWFNet) across three camera lenses and multiple assembly motions.","tokens_in":13340,"tokens_out":5471,"duration_ms":56884,"significance":"If the claims are substantiated, this would be a notable contribution: a single event-camera SHWFS architecture spanning static metrology and high-speed dynamic sensing, with clear advantages in dynamic range and throughput over frame-based approaches. The principle is physically plausible and the paper is generally well written. The throughput definition is transparent, and the inclusion of a fine-tuned passive baseline is a responsible experimental choice. The main weakness is that the dynamic accuracy claim (kilohertz-rate tracking) is validated only against a 30 fps frame-based proxy ground truth that cannot observe the high-frequency motions the sensor is designed to resolve. The static DR claim, meanwhile, is a theoretical bound presented in the abstract without qualification. Both issues are load-bearing for the central contributions.","major_comments":[{"comment":"The dynamic centroid error is computed against a 30 fps frame camera with 20 ms exposure. This reference averages spot motion over 10 modulation cycles at 500 Hz (and 20 at 1 kHz), so all content above ~15 Hz is aliased or lost. Errors are evaluated 'at GT timestamps,' i.e., they quantify agreement with the low-frequency trend of the centroid, not the instantaneous tracking error. The sub-millisecond fluctuations shown qualitatively in Fig. 4c are never quantitatively checked. The mean error being close to the 0.14 px calibration floor is consistent with both sensors agreeing on the low-frequency centroid, not with accurate high-frequency tracking. This is a load-bearing gap: the central claim of kilohertz-rate centroid tracking is not validated. Please add a controlled high-speed experiment (e.g., a spot oscillating at known frequencies/amplitudes with an independent fast reference, or","section":"Sec. 3B, 3C; Figs. 4c–4d, 5e; Table 1"},{"comment":"The static wavefront RMSE reductions (59% and 70%) are based on single acquisitions at two representative sub-apertures, with no repeats or statistical uncertainty. A single realization cannot establish reliability. Please provide multiple acquisitions, error bars, or a bootstrap analysis, and state the number of independent measurements. Additionally, the abstract says 'reaching a dynamic range of 260 dB' without qualification; the body correctly calls it 'theoretical.' Since 260 dB is not directly measured, the abstract should be corrected to avoid implying experimental verification, and the text should state explicitly what intensity range was actually recovered.","section":"Sec. 3A; Fig. 3; abstract"},{"comment":"The intensity-reconstruction model is adopted from [18] without re-verification on the EVK4 sensor used here. Although Fig. 2c–d show latency histograms consistent with the model, a quantitative validation of Eq. (2) using known spot intensities would strengthen the static reconstruction claims. If the model is already well established, please at least discuss its assumptions and any expected deviations for this sensor.","section":"Sec. 2A, Eq. (2)"}],"minor_comments":[{"comment":"State the number of frames used in the 18-second sequence and the number of centroid comparisons that contribute to the error histogram (Fig. 4d).","section":"Sec. 3B"},{"comment":"The sentence 'This exceeds the 500 Hz sensing rate by more than a factor of seven' compares a throughput in samples/s to a modulation frequency in Hz. Please clarify the definition of 'sensing rate' and ensure the units are consistent.","section":"Sec. 3B, Table 2"},{"comment":"The turbulence experiment uses the same 30 fps frame-based proxy ground truth, so the same aliasing limitation applies. Please either address this with a cross-reference or add an explicit limitation statement.","section":"Sec. 3C"},{"comment":"The parameters δ, W, and Tmin are user-defined. Report the values used in the experiments and, ideally, a brief sensitivity analysis.","section":"Sec. 5B, Eqs. (3)–(5)"},{"comment":"For the static comparison, the 500 ms frame-based acquisition used as ground truth is described only briefly. Please provide details of how this GT was computed and how its noise is estimated.","section":"Fig. 3"}],"recommendation":"major_revision","confidential_remarks":"The main issue is the dynamic ground-truth validation. The paper currently claims kilohertz-rate tracking with sub-pixel accuracy, but the only quantitative check is against a 30 Hz, 20 ms-exposure frame camera that cannot resolve the high-frequency motions of interest. The fact that the mean error is near the calibration floor suggests the comparison measures common low-frequency agreement, not high-frequency accuracy. A controlled high-frequency experiment or clearly tempered claims are needed. The static 260 dB claim in the abstract should also be qualified. The framework itself is novel and the throughput analysis is solid; these issues are addressable within the scope of a major revision."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Yuhan Bao et al. have a genuinely new idea: instead of relying on motion-triggered events, they drive the illumination with a periodic waveform and use event timing to encode focal-spot intensity. That lets a single event-camera SHWFS do static one-shot wavefront reconstruction at high dynamic range and dynamic centroid tracking at kilohertz rates. The dual sine/square modulation modes and the Polarity Flip Filter are sensible and, as far as I know, new relative to the passive event-based SHWFS literature. The hardware description is clear, the CPU throughput numbers are transparent, and the comparison with EBWFNet, including a fine-tuned version, is fair. The authors also state limitations (LED power, active-sensing restriction) in the discussion. Credit where it's due: this is a solid within-subfield contribution and the static results in dim and bright sub-apertures look compelling, with 59–70% RMSE reductions versus frame-based SHWFS.\n\nThe soft spot is the dynamic validation. The 'ground truth' for the 500 Hz and 1 kHz centroid tracking is a 30 fps frame camera with 20 ms exposure. That exposure averages over 10–20 modulation cycles, and 30 Hz sampling aliases anything above 15 Hz. The errors are computed at the frame timestamps, so they measure how well the event-based centroids agree with a slowly varying average, not how accurately they track instantaneous spot motion. The sub-millisecond oscillations shown in Fig. 4c are qualitative only. The mean error of 0.16 px being close to the 0.14-pixel calibration floor is consistent with both sensors simply agreeing on the low-frequency centroid. So the paper's central claim—kilohertz-rate accurate tracking—is not actually demonstrated by the data. That's a load-bearing gap, not a nitpick.\n\nOther issues are smaller. The 260 dB dynamic range is a theoretical bound derived from the authors' earlier temporal-mapping model, not a measured value. Static RMSEs are single acquisitions with no repeats, so no variance. The PFF tolerance, sliding-window length, and minimum-pair threshold are not disclosed. Data and code are promised but not currently available.\n\nThis is a paper worth engaging with, not dismissing. The active-modulation idea is good and the static half is fairly solid. The authors need to validate the dynamic half against something that can actually see high-frequency motion—a fast photodetector, a calibrated piezo with known trajectory, or at least a synthetic event stream with a known ground truth. If they can close that gap, the contribution is real. I'd send it to peer review and ask for that additional validation, plus the parameter values and data/code.","headline":"A clever active-modulation scheme that plausibly extends neuromorphic SHWFS to static scenes, but the kilohertz tracking accuracy is validated only against a 30 Hz frame reference—so the headline dynamic claim is not yet supported.","tokens_in":13852,"tokens_out":3232,"would_cite":true,"duration_ms":33223,"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":"By periodically modulating the illumination, a Shack–Hartmann sensor built around an event camera can perform one-shot high-dynamic-range static wavefront reconstruction and kilohertz-rate dynamic tracking on the same hardware.","keywords":["Shack-Hartmann wavefront sensing","event-based vision","neuromorphic sensing","temporal modulation","dynamic range","wavefront reconstruction","centroid tracking","optical metrology"],"falsifier":"A controlled experiment with a spot moving in a known trajectory at sub-millisecond timescales, captured simultaneously by a 1 kHz frame-based camera and the EvTem-SHWFS system, would reveal whether the sub-pixel centroid errors reflect true tracking accuracy; if the 1 kHz camera disagrees with the event-based centroids by more than the reported 0.26 pixels, the dynamic claim is falsified.","tokens_in":12920,"feed_emoji":"⚡","tokens_out":3540,"duration_ms":34347,"temperature":0.7,"pith_summary":"The paper proposes EvTem-SHWFS, a neuromorphic Shack–Hartmann wavefront sensor in which the light source is actively modulated so that focal-spot intensity is encoded into the timing of asynchronous events rather than into pixel brightness. The authors argue that this single architecture resolves two long-standing limits of frame-based wavefront sensors: intensity dynamic range (theoretical 260 dB at 20 Hz, versus ~60 dB for cameras) and the spatial–temporal resolution trade-off (centroid tracking at 500 Hz–1 kHz with microsecond latency). In static metrology, they report 59–70% lower wavefront RMSE than a frame-based sensor under strongly non-uniform illumination; in dynamic scenarios, they report centroid localization errors of 0.18–0.26 pixels, far below a passive event-based baseline, with per-sub-aperture CPU throughput above 420 kHz.","feed_headline":"Modulated light gives wavefront sensing a 260 dB range","feed_subtitle":"A neuromorphic Shack-Hartmann sensor swaps brightness for event timing, cutting wavefront error and tracking centroids at kilohertz rates on","key_machinery":"The load-bearing element is the latency–intensity relationship ρ(x,y) = Qth / ∫_0^{Δt*} S(t) dt, which encodes the static intensity coefficient of the focal spot into the initial positive event latency. The modulation waveform and frequency set the achievable dynamic range: low-frequency sine modulation maximizes it (theoretical 260 dB), while high-frequency square modulation produces temporally compact alternating-polarity event pairs localized to the bright lobe. A Polarity Flip Filter accepts only events whose polarity alternates with a period matching half the modulation cycle, and a sliding-window arithmetic mean yields the centroid.","core_discovery":"The central claim is that temporal modulation converts the focal-spot intensity distribution into a deterministic event-latency map, and that this map serves both operating modes. Under 20 Hz sine modulation, the latency of the initial positive event at each pixel is inversely proportional to the local intensity, so one acquisition recovers full focal-spot energy distributions across a 260 dB dynamic range, enabling moment-based wavefront reconstruction. Under 500 Hz–1 kHz square modulation, alternating-polarity event pairs are generated within each focal spot; a Polarity Flip Filter isolates them and a sliding-window centroid computation tracks the spot at kilohertz rate. The paper shows bo","pith_inferences":["The dynamic accuracy numbers rest on a 30 Hz frame-based reference; if the true spot motion contains high-frequency components that alias, the reported errors could be optimistic. A comparison against a fast camera or a calibrated high-speed spot generator would settle this.","The framework's dependence on active illumination modulation means it applies to active sensing scenarios; extending modulation to a conjugate plane inside the optical system, as the authors suggest, would broaden it to passive scenes.","The system's kilohertz capability is bounded by available photon flux; with a brighter source, the same method might push beyond 1 kHz, since the modulation frequency itself is not otherwise limited."],"forward_implications":["If the claims hold, a single sensor can replace both a high-dynamic-range static wavefront sensor and a high-speed dynamic wavefront sensor, simplifying optical metrology and alignment systems.","The 420,737 Hz per-sub-aperture CPU throughput means kilohertz wavefront sensing no longer requires FPGA or GPU acceleration, potentially lowering cost and latency in closed-loop alignment.","The 260 dB theoretical dynamic range would let one-shot wavefront measurements span intensity variations of nine orders of magnitude, avoiding multiple exposures under non-uniform illumination.","Event-native processing (directly on asynchronous events, without frame construction) removes the event-to-volume conversion bottleneck that dominates the latency of network-based approaches."],"fun_headline_variants":["Neuromorphic wavefront sensor hits 260 dB dynamic range","Kilohertz wavefront tracking on a megapixel neuromorphic sensor","Modulated light gives wavefront sensing 260 dB dynamic range","Event-based wavefront sensing cuts error by up to 70%","One-shot wavefront acquisition with 260 dB dynamic range"],"cache_read_input_tokens":2304,"weakest_assumption_plain":"The dynamic experiments assume that 30 fps frame-based centroids, registered with 0.14-pixel reprojection error, are a correct reference for spot motion that the system is claimed to resolve at 1 kHz; any motion between those 30 Hz samples is invisible to the ground truth.","fun_headline_variants_meta":{"raw":{"variants":["Neuromorphic wavefront sensor hits 260 dB dynamic range","Kilohertz wavefront tracking on a megapixel neuromorphic sensor","Modulated light gives wavefront sensing 260 dB dynamic range","Event-based wavefront sensing cuts error by up to 70%","One-shot wavefront acquisition with 260 dB dynamic range"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000639,"raw_usage":{"total_tokens":2801,"prompt_tokens":786,"completion_tokens":2015,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":530,"completion_tokens_details":{"reasoning_tokens":1941}},"tokens_in":530,"tokens_out":2015,"duration_ms":13852,"temperature":1.0,"reasoning_tokens":1941,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-01T02:51:59.264347+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"A controlled experiment with a spot moving in a known trajectory at sub-millisecond timescales, captured simultaneously by a 1 kHz frame-based camera and the EvTem-SHWFS system, would reveal whether the sub-pixel centroid errors reflect true tracking accuracy; if the 1 kHz camera disagrees with the event-based centroids by more than the reported 0.26 pixels, the dynamic claim is falsified.","supporting_citations":[],"review_version":1}