{"id":"becc2d89-77d0-45ed-a78e-901f646cb616","arxiv_id":"2502.06116","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":2.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"This review traces event cameras from neuromorphic origins to modern back-illuminated stacked sensors and uses blackbody radiometry to suggest 8 micrometers as a promising infrared operating band.","lead":"This paper reviews event-based vision sensors, cameras that respond to scene changes instead of taking full frames. It summarizes recent advances in back-illuminated and stacked-chip designs and argues that infrared event vision is an underexplored opportunity.","discovery_kind":"review","skeptic_critique":{"model":"deepseek-v4-flash","headline":"IR recommendation in Sec. V is based on absolute photocurrent slope, but EVS responds to fractional contrast; the relevant metric d ln I/dT shrinks with wavelength, so the 8 µm conclusion is unjustified.","rationale":"The review's survey content is useful and largely independent of the infrared estimate, but the one original quantitative conclusion in Section V is load-bearing for the paper's stated claim that infrared event vision is a promising frontier. The reader correctly identified that Eqs. (9)-(10) rely on hand-picked parameters and lack validation; the concern raised here is more fundamental: even accepting those parameters, the comparison metric appears to be the wrong one. An EVS is a temporal-contrast sensor, so temperature sensitivity should be measured by fractional photocurrent change per kelvin, not by the absolute slope of the photocurrent curve. In the idealized blackbody model, the fractional derivative decreases with wavelength, so the printed justification for 8 µm does not follow. The manuscript itself notes that MWIR/LWIR EVS studies are rare and that the calculation is illustrative, which supports treating this as a correctness risk that can be fixed by redoing the analysis with a noise-aware SNR metric or by scaling back the recommendation. The overall CONDITIONAL verdict remains appropriate; no adjustment to the reader's verdict is needed.","tokens_in":15797,"tokens_out":9011,"duration_ms":85617,"concrete_test":"Recompute from Eqs. (9)-(10) both dI_ph/dT and (1/I_ph)dI_ph/dT for central wavelengths 3-12 µm at T = 233, 300, and 500 K, using the stated 30 µm pixel, f/2 optics, unity efficiencies, and 0.02 µm band. If the normalized derivative (1/I_ph)dI_ph/dT peaks below 8 µm at all three temperatures, the 'fastest' claim in Fig. 3(b) is an artifact of using absolute slope. Then rerun the wavelength comparison using a detectability metric that includes shot noise and readout noise—e.g., SNR for a 1 K temperature step versus wavelength—and check whether 8 µm remains optimal. If it does not, the conclusion should be revised or explicitly qualified as a noise-dependent design choice.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The paper's sole original quantitative conclusion is the Section V recommendation that an infrared event vision sensor for ground scenes should have a central wavelength around 8 µm, because 'the photocurrent at 8 µm varying the fastest with temperature' (Fig. 3(b)). The load-bearing problem is that an event-based pixel does not respond to absolute photocurrent or its absolute derivative. It triggers on temporal contrast, i.e., the fractional change |Delta I / I|, so the relevant sensitivity metric is |d ln I_ph / dT|, not |d I_ph / dT|. For blackbody photon flux, d ln I / dT = (hc / lambda k T) * e^x / (e^x - 1) / T with x = hc / (lambda k T), which is monotonically decreasing in lambda. At a 300 K scene, this is roughly 3.2%/K at 5 µm versus 2.0%/K at 8 µm. Thus, within the paper's own idealized model of Eqs. (9)-(10), a 5 µm detector would trigger an event for a smaller temperature step than an 8 µm detector, opposite to the stated basis for the recommendation. The 8 µm conclusion could still be recovered in a noise-aware analysis where the higher absolute signal at 8 µm improves SNR, but the paper presents no such analysis. As written, the recommendation mistakes an absolute derivative for the contrast signal that an EVS actually encodes, so the central quantitative claim is unsupported.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper is a review of event-based vision sensors (EVS), tracing the field from neuromorphic engineering to current devices that combine backside illumination, wafer stacking, and industrial interfaces. It surveys the working principle of the EVS pixel, defines performance metrics (contrast sensitivity, dynamic range, power, latency, readout rate), and discusses the potential of infrared event vision. The authors add an original quantitative estimate: using blackbody radiance and a simple photocurrent model, they recommend an infrared detector with a central wavelength around 8 um for ground-scene EVS because the photocurrent at 8 um is said to vary fastest with temperature.","tokens_in":16032,"tokens_out":6309,"duration_ms":56581,"significance":"If the review's factual content is reliable, the paper provides a useful, up-to-date map of the EVS field, particularly the recent progress in BSI processing, wafer stacking, and readout interfaces, and it serves as a compact entry point for researchers interested in infrared event-based imaging. The manuscript cites recent commercial sensors and research prototypes, includes explicit performance metrics, and gives a clear account of the pixel operation. The sole original quantitative contribution, the 8 um wavelength recommendation, is not supported by the argument as written, because the paper applies an absolute photocurrent derivative where the sensor's own operating principle calls for a fractional (logarithmic) derivative; this may be fixable with an SNR-aware analysis, so the issue is significant but not beyond revision.","major_comments":[{"comment":"The recommendation of an 8 um central wavelength rests on the claim that 'the photocurrent at 8 µm varying the fastest with temperature.' An event-based pixel, however, does not respond to the absolute photocurrent or its absolute derivative; it responds to temporal contrast, i.e., the fractional change |ΔI_ph/I_ph|, as defined in Section IV-A and Eqs. (1)-(2). For blackbody radiation, d ln I_ph/dT = (hc/λkT) * e^x/(e^x-1) / T with x = hc/(λkT), which decreases monotonically with λ. At T = 300 K, this is about 3.2%/K at 5 µm versus 2.0%/K at 8 µm. The paper itself uses this fractional metric when stating that 1 K against 300 K at 5 µm requires 3% contrast sensitivity. Therefore, within the paper's own model, the 5 µm detector would trigger an event for a smaller temperature step than the 8 µm detector, opposite to the stated basis for the recommendation. A noise-aware analysis that accounts for the larger absolute photocurrent (and hence better shot-noise-limited SNR) at longer wavelengths could potentially restore the 8 µm conclusion, but no such analysis is presented.","section":"V, Eqs. (9)-(10), Fig. 3(b)"},{"comment":"The photocurrent estimate is an illustrative calculation based on several idealized assumptions (30 µm pixel pitch, f/2 optics, unity optical transmission and detection efficiency, and a 5±0.01 µm spectral band). The resulting 1-862 pA range has no uncertainty bounds and is not validated against measured infrared detector photocurrents. Since this calculation is the quantitative basis for the central infrared recommendation, please add a sensitivity analysis with respect to pixel pitch, f-number, efficiency, and spectral bandwidth, and state whether the conclusion is robust to realistic parameters.","section":"V, Eqs. (9)-(10)"},{"comment":"There is an internal inconsistency in the treatment of contrast. Section IV-A correctly defines temporal contrast as a fractional change, and Section V uses this metric to compute the 3% and 30% contrast sensitivities for 1 K and 10 K changes at 5 µm. However, the wavelength recommendation is based on the absolute rate of change of photocurrent with temperature. These two criteria lead to different preferred wavelengths, and the manuscript does not reconcile them. Please clarify whether the recommendation is based on contrast sensitivity, SNR, or another figure of merit, and revise the argument accordingly.","section":"IV-A and V"}],"minor_comments":[{"comment":"The conclusion section is numbered VII, but no section VI appears; please renumber the sections sequentially.","section":"General"},{"comment":"Several equations are garbled in the submitted text, with missing operators, limits, and subscripts, making them difficult to verify; please ensure the final manuscript has properly typeset equations.","section":"Equations (1)-(10)"},{"comment":"The text refers to 'Table 1' for benchmarks, but no table appears in the supplied manuscript; verify that Table 1 is present with all columns and values.","section":"Section IV"},{"comment":"Figures 1-3 are cited but not visible in the extracted text; confirm that all figures and captions are included.","section":"Figures"},{"comment":"The phrase 'may be the identical choice' should likely read 'may be the ideal choice' or 'may be an optimal choice.'","section":"Section V"},{"comment":"The phrase 'serval contributions' is a typo for 'several contributions.'","section":"Section II"},{"comment":"The sentence fragment 'Compared to visible light.' should be joined with the adjacent sentence.","section":"Section V"}],"recommendation":"major_revision","confidential_remarks":"The survey content appears informative and generally accurate, but the original quantitative section needs substantive rework. The authors should either correct the optical-wavelength recommendation to use the temporal-contrast metric or supply the missing noise/SNR analysis; without that, the paper's main forward-looking claim is unsupported. The many formatting problems in the submitted text (missing equations, missing table, missing figures) may be extraction artifacts, but they should be checked carefully before publication."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Colleague,\n\nThis review of event-based vision sensors is worth reading for the survey half. It pulls the DVS/ATIS/DAVIS lineage, the BSI and wafer-stacking steps, and the industrial interface picture (MIPI, Celex scan readout, Prophesee's on-chip pipeline) into one place. The feature taxonomy—contrast sensitivity, dynamic range, power, latency, readout rate—is sensible, and the device numbers match what I know from the literature. If you need a quick map of where EVS hardware stands, this is a decent starting point.\n\nThe soft spot is Section V. The paper's only original quantitative conclusion is that an 8 µm central wavelength is 'the identical choice' for ground-scene EVS because the absolute photocurrent varies fastest with temperature there. That is the wrong metric. An EVS pixel responds to temporal contrast, i.e. the fractional change in photocurrent, because the front end compresses logarithmically. The relevant figure of merit is d ln I_ph / dT, which decreases with wavelength for blackbody radiation: about 3.2%/K at 5 µm versus 2.0%/K at 8 µm for a 300 K scene. So within the paper's own idealized model, a 5 µm detector would give a larger contrast signal for the same temperature step. The 8 µm recommendation could be rescued with a noise or SNR argument, but the paper does not make one. The text even computes that a 1 K change at 5 µm needs ~3% contrast sensitivity, which sits uneasily with the 8 µm choice.\n\nThere are also manuscript hygiene problems: leftover template text, a missing Section VI, duplicated references, and equations that are garbled in the version I read. That does not change the survey's value, but it needs a cleanup.\n\nThe IR analysis should be corrected—compare fractional contrast, or add an SNR model—or be explicitly labeled as a first-order absolute-photocurrent comparison that ignores the sensor's logarithmic response. As written, I would not rely on the 8 µm recommendation. The survey itself deserves referee time because the field needs a consolidated hardware summary.\n\nRecommendation: send to peer review with heavy revision expected for Section V and the manuscript cleanliness. The review is useful after fixes.","headline":"The survey half of this paper is a useful map of event-vision sensors, but the only original quantitative claim—an 8 µm recommendation for infrared EVS—is based on absolute photocurrent rather than temporal contrast, and is wrong as stated.","tokens_in":16619,"tokens_out":6483,"would_cite":true,"duration_ms":53808,"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":"This review traces how event-based vision sensors matured into practical industrial devices and argues, on the basis of a blackbody photocurrent model, that an infrared detector centered near 8 µm is the best choice for ground-scene event…","keywords":["event camera","backside illumination","industry camera interface","wafer-stacking","infrared imaging","temporal contrast","neuromorphic engineering","dynamic vision sensor"],"falsifier":"Measure, for a set of real or simulated infrared detectors with different cutoff wavelengths, the equivalent photocurrent as a function of blackbody temperature across -40°C to 500°C, and find the wavelength where the photocurrent derivative with temperature is largest; if that wavelength is not near 8 µm, the paper's central infrared recommendation fails.","tokens_in":15562,"feed_emoji":"📷","tokens_out":9483,"duration_ms":76371,"temperature":0.7,"pith_summary":"This review traces how event-based vision sensors, inspired by biological retinas, grew from laboratory prototypes into practical industrial devices. It argues that backside illumination, wafer stacking, and standard industrial interfaces were the key enablers of lower noise, higher resolution, and faster readout. It then turns to the thermal infrared, where event vision is still rare, and uses a blackbody photocurrent model to propose that a detector centered near 8 µm is the best match for ground-scene event imaging. The review matters because it consolidates the field's scattered metrics and designs into one map and gives a first-order wavelength target for anyone building an infrared event camera.","feed_headline":"8 µm is the best infrared wavelength for event cameras","feed_subtitle":"Back-illumination and stacking made event vision practical; the model points to 8 µm for thermal scenes.","key_machinery":"Two mechanisms carry the argument. The first is the logarithmic temporal-contrast pixel core—photodiode, log trans-impedance amplifier, self-timed differential encoder, and threshold comparators—which converts relative intensity change into asynchronous on/off events. The second is the blackbody radiation model of Eqs. (9)–(10), which turns scene temperature into equivalent detector photocurrent $I_{\\mathrm{equ}} = B_L A_d \\eta_{\\mathrm{optics}}\\eta_{\\mathrm{det}}/(4(f/\\#)^2)$; because the photocurrent at 8 µm changes fastest with temperature, the model singles out that wavelength for ground-scene event imaging.","core_discovery":"The paper's central claim is that event-based vision has reached industrial viability through three technology shifts: backside illumination raising fill factor and quantum efficiency, wafer stacking enabling on-chip processing and high-resolution hybrid sensors, and industrial interfaces such as MIPI removing readout bottlenecks. Extending the same logic to the thermal infrared, it claims that for ground scenes an infrared detector with a central wavelength around 8 µm is the ideal choice for an event sensor, because that is where the equivalent photocurrent changes fastest with scene temperature. Supporting this, it reports that the field now uses metrics like nominal contrast threshold, dynamic energy, and static power, and that infrared event imaging faces a dynamic range of about 70 dB and needs roughly 3% contrast sensitivity to detect a 1 K change at 5 µm against a 300 K background.","pith_inferences":["The 8 µm recommendation is an output of an idealized model; a natural next step is to repeat the calculation for realistic detector parameters (e.g., non-unity quantum efficiency, f/1.2 optics, larger or smaller pixels) and check whether the optimal wavelength moves.","The same temperature-derivative logic could be turned into a general design rule: choose the wavelength where the Planckian derivative peaks for the scene temperature of interest, which for hot industrial or space scenes would shift the optimum away from 8 µm.","The absence of a standardized EVS evaluation protocol suggests an opportunity to convert the paper's metric collection into a common test procedure, which would make future sensor comparisons meaningful."],"forward_implications":["Designers building infrared event cameras for terrestrial scenes would target detectors near 8 µm rather than mid-wave infrared wavelengths, and would use the model to budget pixel size, f-number, and efficiency.","The demonstrated gains from backside illumination and wafer stacking imply that future event sensors can push resolution and readout rates beyond current figures without sacrificing sensitivity.","The compiled evaluation metrics give the field a common language, so new sensors can be compared on nominal contrast threshold, power at defined event rates, and latency rather than on ad hoc specifications.","On-chip processing and hybrid frame-plus-event outputs make event sensors a practical component for edge vision, robotics, and automotive systems rather than a research curiosity.","Infrared event cameras will need front-end designs that cope with large dark currents, since the achievable dynamic range is near 70 dB and detecting a 1 K temperature change at 5 µm against a 300 K background requires roughly 3% contrast sensitivity."],"supporting_citations":[{"why":"Reports the first event sensor with backside illumination and a MIPI interface, achieving 300 Mbps output bandwidth; load-bearing for the claim that industrial interfaces unlocked high readout rates.","marker":"[11]"},{"why":"Describes a three-wafer-stacked hybrid 15-MPixel CIS plus 1-MPixel EVS with 4.6 Gevent/s readout; load-bearing for the wafer-stacking and readout-rate claims.","marker":"[14]"},{"why":"Presents a BSI-CMOS stacked event sensor with on-chip processing, supporting the claim that integration has moved event vision to single-chip systems.","marker":"[21]"},{"why":"Compares frontside- and backside-illuminated versions of the same DAVIS sensor, showing fill factor and quantum efficiency gains; load-bearing for the BSI benefits.","marker":"[55]"},{"why":"Reports an uncooled microbolometer LWIR event sensor and the ~1% contrast needed for 1 K changes against a 310 K background; load-bearing for infrared EVS feasibility.","marker":"[71]"},{"why":"Describes the two-stage capacitive-feedback differencing amplifier used for temporal-contrast infrared sensors and its speed limitations; load-bearing for the claim that IR event front-ends are constrained.","marker":"[72]"},{"why":"Evaluates a photoreceptor circuit for a III-V nBn detector with 5.5 µm cutoff at 130 K, showing dark current limits dynamic range; load-bearing for the infrared challenges.","marker":"[73]"},{"why":"Supplies the infrared scene properties and ~70 dB dynamic range estimate for ground scenes used in the model comparison.","marker":"[70]"}],"fun_headline_variants":["Event cameras find ideal IR wavelength: 8 µm","Back-illumination and stacking make event vision practical","Event vision hits industrial stride with BSI and stacking","Why 8 µm is the sweet spot for thermal event sensors","Event-based vision: from lab to edge with new tech"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The 8 µm recommendation rests on an illustrative photocurrent model that assumes a 30 µm pixel, f/2 optics, unity optical transmission and detection efficiency, and a narrow spectral band at 5 µm; real infrared detectors with different parameters could shift the preferred wavelength.","fun_headline_variants_meta":{"raw":{"variants":["Event cameras find ideal IR wavelength: 8 µm","Back-illumination and stacking make event vision practical","Event vision hits industrial stride with BSI and stacking","Why 8 µm is the sweet spot for thermal event sensors","Event-based vision: from lab to edge with new tech"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000305,"raw_usage":{"total_tokens":1714,"prompt_tokens":870,"completion_tokens":844,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":486,"completion_tokens_details":{"reasoning_tokens":764}},"tokens_in":486,"tokens_out":844,"duration_ms":6794,"temperature":1.0,"reasoning_tokens":764,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-08T16:42:37.521842+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Measure, for a set of real or simulated infrared detectors with different cutoff wavelengths, the equivalent photocurrent as a function of blackbody temperature across -40°C to 500°C, and find the wavelength where the photocurrent derivative with temperature is largest; if that wavelength is not near 8 µm, the paper's central infrared recommendation fails.","supporting_citations":[{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Reports the first event sensor with backside illumination and a MIPI interface, achieving 300 Mbps output bandwidth; load-bearing for the claim that industrial interfaces unlocked high readout rates."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Compares frontside- and backside-illuminated versions of the same DAVIS sensor, showing fill factor and quantum efficiency gains; load-bearing for the BSI benefits."},{"cited_title":"Posch, D","cited_arxiv_id":null,"evidence_quote":"Reports an uncooled microbolometer LWIR event sensor and the ~1% contrast needed for 1 K changes against a 310 K background; load-bearing for infrared EVS feasibility."},{"cited_title":"Posch, D","cited_arxiv_id":null,"evidence_quote":"Describes the two-stage capacitive-feedback differencing amplifier used for temporal-contrast infrared sensors and its speed limitations; load-bearing for the claim that IR event front-ends are constrained."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Evaluates a photoreceptor circuit for a III-V nBn detector with 5.5 µm cutoff at 130 K, showing dark current limits dynamic range; load-bearing for the infrared challenges."},{"cited_title":"A Survey on Infrared Image and Video Sets","cited_arxiv_id":"2203.08581","evidence_quote":"Supplies the infrared scene properties and ~70 dB dynamic range estimate for ground scenes used in the model comparison."}],"review_version":1}