{"id":"2d89648f-a491-41e1-a933-de6539a1913b","arxiv_id":"2411.17313","paper_version":2,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":7.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"Event Ellipsometer reconstructs normalized Mueller-matrix videos at 30 fps from event-camera streams produced by fast-rotating quarter-wave plates.","lead":"Event Ellipsometer is a new camera system that records how surfaces change the polarization of light, as a Mueller-matrix video, at 30 frames per second using an event camera and two rapidly rotating quarter-wave plates. A generalist might care because it extends ellipsometry, historically a slow static-lab technique, to moving scenes such as faces, transparent tape, and high-dynamic-range objects.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The 30fps dynamic Mueller-matrix claim rests on ignoring intra-frame motion; the paper's own Supplement 5.3 admits motion-dominated events corrupt reconstruction, so the demonstrated scope is quasi-static, not general dynamic scenes.","rationale":"The reader's weakest-assumption analysis correctly identifies the intra-frame stationarity of M as the load-bearing premise for the 30fps dynamic-scene claim. I considered the alternative concern raised in the reader's rationale, the dimensional inconsistency in Eq. (5) where derivative terms appear to omit the factor ω. That issue affects reproducibility, but it is more plausibly an unstated rescaling or typo because the authors report small MSEs on known optical elements, which would not occur if the printed equation were implemented literally. The motion concern, by contrast, is acknowledged by the authors themselves in Supplement 5.3 and directly undermines the central claim's scope: if event generation is dominated by object motion, the estimated M is not the material's Mueller matrix. The paper provides no quantitative motion sensitivity analysis, and the hair scene exhibits exactly the predicted artifacts. Therefore the existing CONDITIONAL verdict remains appropriate; the paper should add a controlled motion experiment and a quantitative event-ratio analysis before the dynamic-scene claim is accepted at face value. The proposed test would settle whether the concern lands by measuring how reconstruction error scales with intra-frame motion and event-source dominance.","tokens_in":18269,"tokens_out":10828,"duration_ms":103785,"concrete_test":"Mount a known birefringent sample, such as a linear polarizer or quarter-wave plate, on a motorized translation or rotation stage and capture events while the stage moves at controlled speeds, with per-frame displacements ranging from 0.05 to 5 pixels and motion frequencies from 0.5 to 15 Hz. Reconstruct the normalized Mueller matrix for each 33ms frame and compute the MSE against the static ground-truth matrix, plotting MSE versus displacement per frame and versus the ratio of motion-induced to polarization-induced events. If the MSE exceeds 0.1 (roughly 2x the reported 0.045) at sub-pixel displacements, or if error rises sharply when the motion-to-polarization event ratio approaches 0.5, then the intra-frame stationarity assumption fails for realistic dynamic scenes and the 'dynamic video' claim must be explicitly restricted to quasi-static scenes.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim of Mueller-matrix video at 30fps for dynamic scenes requires that within each 33ms frame the scene Mueller matrix M is time-invariant, as assumed in Eq. (4). Any motion of the target introduces an unmodeled dM/dt term in the log-intensity derivative, and moving edges generate events that are interpreted by Eq. (7) as if they arose purely from the rotating QWPs. The supplement (Sec. 5.3) explicitly concedes that reconstruction accuracy degrades when event generation is dominated by object motion and documents motion artifacts in the hair scene, yet the paper provides no quantitative evaluation of reconstruction error versus motion amplitude or versus the motion-to-polarization event ratio. The demonstrations of facial expression and hair motion therefore do not establish that the system recovers the correct Mueller matrix under realistic dynamics; they show only that plausible images can be produced. Without a controlled motion study, the headline extension to dynamic scenes is not supported beyond quasi-static motion.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"This paper presents Event Ellipsometer, a Mueller-matrix video imaging system that pairs a Prophesee EVK4 event camera with two fast-rotating quarter-wave plates (QWPs), one in the illumination path and one in the detection path, plus fixed linear polarizers. The authors derive a Stokes-Mueller image formation model in which the logarithmic intensity derivative is expressed as a ratio of linear forms in the vectorized Mueller matrix (Eqs. (4)-(7)), and they formulate reconstruction as a per-pixel weighted least-squares SVD with Cloude physical-validity filtering followed by spatio-temporal propagation and refinement. A one-time calibration estimates the per-pixel contrast threshold and the QWP offset angles. Experiments include synthetic validation, real measurements of known optical elements and a metal plate, and demonstrations on photoelasticity, transparent tape detection, dynamic human face/hair, and HDR scenes. The paper claims Mueller-matrix video capture at 30 fps, i.e., a 33 ms frame duration, and reports mean-squared errors on the order of 0.013-0.020 for known elements.","tokens_in":18472,"tokens_out":9423,"duration_ms":77909,"significance":"The contribution is potentially significant: if the 30 fps capture claim holds, it would extend Mueller-matrix ellipsometry from static samples requiring minutes of acquisition to dynamic scenes, opening new applications in material inspection, photoelasticity, and human capture. The paper provides a substantial amount of engineering detail (hardware prototype, part list, motor synchronization, calibration procedures) that would aid reproducibility, and it clearly identifies limitations such as the non-real-time reconstruction pipeline. The synthetic validation and the comparison with a frame-based method in the supplement are useful. However, the significance hinges on two points: the correctness of the derivative in Eq. (5), which as printed is wrong, and the validity of the dynamic-scene claim, which is only weakly supported.","major_comments":[{"comment":"The printed expression for dA_t/dt omits the angular velocity factors from every nonzero entry. For example, the derivative of α1^2 is -4ω α1α2, but the paper lists -4α1α2; similarly the α3/α4 terms miss the factor 5ω. With ω = 30π rad/s, the derivative term in Eq. (7) is consequently smaller by two orders of magnitude than it should be, and the units of the B_tk matrix become inconsistent. Because Eq. (7) is the basis of the entire reconstruction (Section 5) and the reported experimental MSE values, the results cannot be reproduced from the equations as written. The authors must correct Eq. (5) (and any dependent equations in the supplement) and confirm whether the implementation used the corrected formula.","section":"Section 4, Eq. (5)"},{"comment":"The paper's headline claim is Mueller-matrix video imaging at 30 fps for dynamic scenes, but the method assumes that the scene Mueller matrix is constant within each 33 ms frame (Eq. (4)) and that event generation is dominated by the rotating QWPs. Supplement 5.3 explicitly states that reconstruction accuracy degrades when object motion dominates event generation and reports motion artifacts in the human hair scene. Yet the paper offers no quantitative evaluation of how reconstruction error depends on motion amplitude, motion speed, or the ratio of motion-induced to polarization-induced events. The dynamic demonstrations (facial expression change, head rotation) therefore show that plausible images can be produced, not that correct Mueller matrices are recovered under the demonstrated dynamics. Please add a controlled motion experiment (e.g., a known Mueller-matrix target undergoing translations or rotations with varying speed) or substantially weaken the dynamic-scene claim to quasi-static scenes.","section":"Sections 3 and 7; Supplement 5.3"}],"minor_comments":[{"comment":"The introduction states 'achieving a mean-squared error of 0.045 for materials with known Mueller matrices,' but Figure 5(a) reports MSEs of 0.016, 0.015, 0.013, 0.020, and 0.020; please clarify which number is the reported MSE and harmonize the text.","section":"Introduction and Section 7"},{"comment":"The depolarization factor α is set to 0.8 for the reference QWP model without a sensitivity analysis; please report how the calibration error varies with α or justify the choice.","section":"Supplement 3.5"},{"comment":"The polarization preservation formula reads ρ = 1/3(M∆,11 + M∆,20 + M∆,11), which appears to contain a typo (likely M∆,22 instead of M∆,20); please correct.","section":"Supplement 4.6, Eq. (17)"},{"comment":"The notation '1' for the all-ones matrix and N for the Gaussian perturbation should be defined more explicitly to avoid confusion with the identity matrix.","section":"Section 5.2, Eq. (12)"},{"comment":"The timeline schematic is hard to read at the current size; the event markers and the log-intensity trace should be enlarged for legibility.","section":"Figure 2(b)"},{"comment":"The statement that existing event-based vision methods cannot capture full polarization reflectance properties as a Mueller-matrix image would benefit from a citation to the most recent works in event-based polarimetry.","section":"Section 2"}],"recommendation":"major_revision","confidential_remarks":"To the editor: The error in Eq. (5) is most likely a typographical omission of ω, but it must be fixed before publication because the reconstruction equations are dimensionally inconsistent as printed. The dynamic-scene claim is the weakest part of the paper; the self-admitted limitation in Supplement 5.3 should be either experimentally addressed or clearly stated in the abstract. If the authors can correct the equation and provide a controlled motion study, the paper would be a solid systems contribution; otherwise the claims should be scaled back."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Prashant, this one is worth a look. The paper combines a dual rotating quarter-wave plate ellipsometer with an event camera to reconstruct normalized Mueller matrices at 30 fps per pixel. That combination is new: prior event-based polarimetry only handled linear polarization with a single rotating polarizer, and prior dual-rotating-retarder systems took minutes. The image formation model, the calibration of contrast threshold and QWP offsets, and the two-stage reconstruction with Cloude filtering and spatiotemporal propagation are all clearly described. The real-world validation on known optical elements (MSE 0.013–0.020) and the comparison against a 5-minute frame-based acquisition are convincing. I also give them credit for shipping a working prototype and for being upfront in the supplement about wavelength dependency, temporal artifacts, and motion artifacts.\n\nNow the soft spots. The printed Eq. (5) omits the angular velocity ω from every term of dA_t/dt. Since ω=30π, the derivative term in the reconstruction equation is two orders of magnitude too small if taken literally; the method could not work as written. I assume it's a typo—their experiments are too good for the printed math to be what they ran—but it must be fixed before anyone can reproduce the derivation. The second issue is the dynamic-scene claim. The supplement's Sec. 5.3 admits reconstruction degrades when event generation is dominated by object motion, and the face/hair demos are qualitative. There is no controlled study of reconstruction error vs. motion amplitude or vs. motion-to-polarization event ratio. So the title's 'Mueller-matrix video' is really 'quasi-static scenes sampled at 30 Hz,' which is still useful but not the same as general dynamic ellipsometry. Minor: the output is normalized by M00 (they acknowledge), and no code/data are released, which limits reproducibility.\n\nOverall the central idea is sound, the experiments are plausible, and the limitations section is honest. This deserves serious peer review, but the referee should demand a corrected Eq. (5) and a motion-robustness experiment. I'd bring it to a reading group if you're working in computational imaging.","headline":"Clever combination of event cameras and rotating retarders for 30 fps Mueller-matrix imaging; the printed derivative drops the ω factor, and 'dynamic' really means quasi-static within each frame.","tokens_in":19022,"tokens_out":3274,"would_cite":false,"duration_ms":32675,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"Rotating quarter-wave plates and an event camera estimate the normalized Mueller matrix per pixel at 30 fps.","keywords":["event camera","Mueller matrix","ellipsometry","polarimetric imaging","dynamic scene capture","rotating quarter-wave plate","high dynamic range imaging","photoelasticity"],"falsifier":"Move a known polarimetric target, such as a linear polarizer, fast enough that motion-induced events outnumber modulation-induced events, reconstruct the Mueller matrix from that frame, and check whether the mean-squared error against the known ground truth exceeds the 0.045 reported for static scenes.","tokens_in":1300,"feed_emoji":"🌀","tokens_out":2108,"duration_ms":119215,"temperature":0.7,"pith_summary":"This paper's central claim is that ellipsometry, the measurement of a material's 4x4 Mueller matrix, can be taken from minutes-long static captures down to 33 ms video frames. The system rotates two quarter-wave plates continuously, one in front of a light source and one in front of an event camera, so that polarization modulation alone drives the camera's events. A derivation relates the time differences between consecutive events to the normalized Mueller matrix of each pixel, and a two-stage estimator with physical-validity filtering reconstructs per-pixel matrices at 30 fps. Experiments report a mean-squared error of 0.045 on known polarimetric samples and show working Mueller-matrix videos of human faces, hair, gelatine under stress, and transparent tape.","feed_headline":"Rotating wave plates capture Mueller-matrix video at 30 fps","feed_subtitle":"Ellipsometry once took minutes per image; now dynamic scenes can be measured in 33 ms frames.","key_machinery":"The central object is the dual-rotating-retarder ellipsometer arrangement: a linear polarizer and quarter-wave plate in front of the light source, and another quarter-wave plate and linear polarizer in front of the event camera, with the camera-side wave plate rotating five times faster than the source-side one. This continuous rotation encodes the sixteen Mueller-matrix elements into the temporal intensity profile seen by the event camera. The key identity is the event-camera threshold relation $\\partial \\log I_t / \\partial t = p_k C / \\Delta t_k$, which converts measured inter-event time differences into linear equations in the unknown Mueller matrix, so the camera's asynchronous events are used directly as the measurement signal.","core_discovery":"The paper establishes that the asynchronous event stream of a dual rotating quarter-wave-plate ellipsometer is sufficient to estimate the normalized Mueller matrix per pixel. In the image formation model $I_t = A_t \\hat{M}$, $\\hat{M}$ is the vectorized Mueller matrix and $A_t$ captures the time-varying polarimetric modulation; because the event camera responds to $\\partial \\log I_t / \\partial t$, the observed time differences $\\Delta t_k$ of events satisfy $B_{t_k} \\hat{M} = 0$ for a known system matrix $B$. Stacking these constraints over events in one frame gives a weighted least-squares problem, solved by SVD, then refined by Cloude's physical-validity projection and a spatiotemporal propagation scheme. The result is a 30 fps Mueller-matrix video for non-planar, dynamic, and high-dynamic-range scenes, reconstructed from 33 ms of events per frame rather than many minutes of frame-based capture.","pith_inferences":["Pairing this event camera with a conventional intensity sensor could recover the absolute, unnormalized Mueller matrix, since the missing scale is exactly the intensity the event camera discards.","The fixed 33 ms frame boundary could become an adaptive window that detects when motion-induced events dominate, yielding a motion-aware confidence per pixel.","Adding a narrow bandpass filter would remove the wavelength-dependent mixing of the white LED and monochrome sensor, at the cost of lower light levels.","A GPU implementation of the reconstruction would likely close the gap between 30 fps capture and the current offline reconstruction time, making live ellipsometric video feasible."],"forward_implications":["Mueller-matrix imaging extends from static samples to dynamic scenes, because capture time per frame drops from minutes to 33 ms.","Non-planar objects can be measured without sacrificing sensor spatial resolution, unlike snapshot metasurface-based approaches.","High-dynamic-range scenes are captured in a single event stream, without bracketed exposures.","Demonstrated applications include photoelastic stress visualization in transparent materials, transparent tape detection, and polarimetric capture of human faces and hair.","The reconstructed quantity is the normalized Mueller matrix $M/M_{00}$, and recovering absolute scale is left as future work."],"supporting_citations":[{"why":"establishes the dual rotating retarder architecture for Mueller-matrix polarimetry that the optical setup is based on.","marker":"[3]"},{"why":"provides the optimization of the dual-rotating-retarder polarimeter including the 5:1 rotation speed ratio used here.","marker":"[44]"},{"why":"supplies the event-camera model relating inter-event time differences to the log-intensity derivative, used in the image formation equation.","marker":"[21]"},{"why":"supplies Cloude's physical-validity filter used to project reconstructed Mueller matrices onto physically realizable space.","marker":"[16]"},{"why":"provides the frame-based polarimetric BRDF acquisition baseline and known material Mueller matrices used for validation.","marker":"[8]"},{"why":"supplies the event simulator used to convert rendered intensity frames into synthetic event streams for evaluation.","marker":"[34]"},{"why":"is the single-shot Mueller-matrix imaging method compared against, which trades spatial resolution and planar scene assumptions.","marker":"[49]"}],"fun_headline_variants":["Event camera yields 30fps Mueller-matrix video from rotating wave plates","Mueller-matrix video at 30fps via event-based ellipsometry","Rotating wave plates and event camera capture dynamic Mueller matrices","Dynamic scenes now measurable: 30fps Mueller-matrix video","Ellipsometry for video: 30fps Mueller-matrix from event data"],"cache_read_input_tokens":21248,"weakest_assumption_plain":"The load-bearing premise is that within each 33 ms frame the scene's Mueller matrix is constant and that the rotating wave plates, not object motion, dominate event generation.","fun_headline_variants_meta":{"raw":{"variants":["Event camera yields 30fps Mueller-matrix video from rotating wave plates","Mueller-matrix video at 30fps via event-based ellipsometry","Rotating wave plates and event camera capture dynamic Mueller matrices","Dynamic scenes now measurable: 30fps Mueller-matrix video","Ellipsometry for video: 30fps Mueller-matrix from event data"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.0007,"raw_usage":{"total_tokens":3130,"prompt_tokens":883,"completion_tokens":2247,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":499,"completion_tokens_details":{"reasoning_tokens":2150}},"tokens_in":499,"tokens_out":2247,"duration_ms":13133,"temperature":1.0,"reasoning_tokens":2150,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-12T12:17:16.344303+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Move a known polarimetric target, such as a linear polarizer, fast enough that motion-induced events outnumber modulation-induced events, reconstruct the Mueller matrix from that frame, and check whether the mean-squared error against the known ground truth exceeds the 0.045 reported for static scenes.","supporting_citations":[{"cited_title":"Application of ellipsometry techniques to biological materials","cited_arxiv_id":null,"evidence_quote":"establishes the dual rotating retarder architecture for Mueller-matrix polarimetry that the optical setup is based on."},{"cited_title":"Instant dehazing of images using polarization","cited_arxiv_id":null,"evidence_quote":"provides the optimization of the dual-rotating-retarder polarimeter including the 5:1 rotation speed ratio used here."},{"cited_title":"Acquiring the reflectance field of a human face","cited_arxiv_id":null,"evidence_quote":"supplies the event-camera model relating inter-event time differences to the log-intensity derivative, used in the image formation equation."},{"cited_title":"Polarization and phase-shifting for 3d scanning of translucent objects","cited_arxiv_id":null,"evidence_quote":"supplies Cloude's physical-validity filter used to project reconstructed Mueller matrices onto physically realizable space."},{"cited_title":"All-photon polarimetric time-of-flight imaging","cited_arxiv_id":null,"evidence_quote":"provides the frame-based polarimetric BRDF acquisition baseline and known material Mueller matrices used for validation."},{"cited_title":"Deep polarization cues for transparent object segmentation","cited_arxiv_id":null,"evidence_quote":"supplies the event simulator used to convert rendered intensity frames into synthetic event streams for evaluation."},{"cited_title":"The mechanical and photoelastic properties of 3d printable stress-visualized materials","cited_arxiv_id":null,"evidence_quote":"is the single-shot Mueller-matrix imaging method compared against, which trades spatial resolution and planar scene assumptions."}],"review_version":1}