{"id":"fa979d13-0ef3-443d-a5de-7b72e228b00c","arxiv_id":"2504.19253","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":5,"one_line_summary":"A controlled benchmark shows event cameras lose accuracy in fast, cluttered scenes while the Tianmouc sensor stays stable across speeds.","lead":"This paper measures how two types of brain-inspired vision sensors hold up in fast robotic perception, using a turntable and printed patterns. It finds event cameras degrade in fast, cluttered scenes, while the Tianmouc chip stays stable across speeds, a result that could guide sensor choice in drones and robots.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Algorithm asymmetry is load-bearing: Tianmouc is scored with the authors' tuned pipelines while EVS use fixed third-party tools, so the hardware-level conclusion may be an artifact of software maturity.","rationale":"The reader's weakest assumption identifies exactly the load-bearing concern I would raise: sensor performance is confounded with algorithmic pipeline choice. The paper's central claim is causal and hardware-level, so it requires that conclusions be robust to reasonable algorithm choices, or at least that the asymmetry be tested. No such test appears. This is not an external quibble: Section VI explicitly defers 'more representative algorithms' to future work, and Section V.B uses different temporal integration strategies for Tianmouc versus event cameras, which alone can produce the observed high-speed degradation in event flow. The absence of raw data, code, and error bars compounds the issue but is secondary; even with perfect reproducibility, the algorithm confound remains. I do not recommend rejecting the paper: the standardized platform, calibration method, and comparative data are valuable, and the hardware-level advantage may survive a crossover test. But the current evidence supports a conditional verdict requiring algorithm-crossover results and data/code release, not an unconditional accept. My recommendation is therefore UNCHANGED relative to the reader's CONDITIONAL verdict.","tokens_in":11773,"tokens_out":3622,"duration_ms":44641,"concrete_test":"Run a crossover benchmark on the recorded sequences: for optical flow, apply the same estimator (e.g., EV-FlowNet or a recent event-flow network) to both raw event streams from DVXplorer/DAVIS346 and event-like streams generated by thresholding Tianmouc AOP-TD differences, using identical angular integration windows (1.5 and 15 degrees) for both sensors; for corner detection, run Shi-Tomasi on E2VID reconstructions from event data and on Poisson-blended Tianmouc AOP-SD reconstructions. If Tianmouc no longer dominates under matched algorithms and windows, the central hardware-attribution claim is unsupported.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim attributes EVS degradation and Tianmouc robustness to sensor-level mechanisms (pixel bandwidth variation, event rate saturation, spatiotemporal-gradient sampling). For that attribution to hold, the post-processing algorithms paired with each sensor must not be the source of the difference. The protocol violates this condition. In Section V.A, Tianmouc corners come from Poisson-blended AOP-SD reconstructions processed with Shi-Tomasi, while event corners come from ARC* plus a 3-pixel filter. In Section V.B, Tianmouc optical flow uses 'the approach described in [2]' with dynamic frame intervals aligned to >1 degree of rotation, whereas event flow uses EV-FlowNet with fixed 1.5 and 15 degree integration windows. Section IV.E similarly pairs E2VID with event data but the authors' own reconstruction/interpolation code with Tianmouc. The 15 degree fixed window at high speed can blur the event flow estimate, reproducing the observed degradation without invoking bus saturation. The paper's own Discussion (Section VI) admits event algorithms are advancing and that 'more representative algorithms' are future work, and also states Tianmouc's software is immature, but no sensitivity analysis or crossover experiment is provided. Because the headline conclusion is about sensor selection, this asymmetry is load-bearing: if Tianmouc's advantage shrinks or disappears under matched algorithms, the central hardware-level claim is not established.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes a quantitative evaluation framework for comparing brain-inspired vision sensors in high-speed robotic perception. Three sensors are tested under controlled turntable rotation (50–3000 rpm) and two illumination levels: the event-based DVXplorer, the RGB-hybrid DAVIS346, and the primitive-based Tianmouc sensor. The evaluation covers imaging quality (edge thickness, structural-information metrics, BRISQUE on reconstructed frames) and functional tasks (corner detection, optical-flow-based angular-velocity estimation). The main claim is that EVS performs well in sparse high-speed scenes and moderately fast complex scenes, but degrades in high-speed cluttered scenes due to pixel-level bandwidth variation and event-rate saturation, whereas Tianmouc maintains consistent performance across all tested speeds and scene complexities due to its global spatiotemporal-gradient sampling.","tokens_in":12071,"tokens_out":3351,"duration_ms":38215,"significance":"If the central claim is established, the paper provides a useful, scenario-dependent sensor-selection guide for high-speed robotics and introduces a reusable evaluation methodology that bridges different BVS data modalities. The controlled turntable setup with ground-truth rotation speeds, the motion-compensated event calibration, and the use of multiple complementary metrics are genuine strengths. The paper is also candid in Section VI about the immaturity of Tianmouc software and the rapid development of event algorithms. However, the headline conclusion is sensor-level, and the current protocol does not adequately separate sensor hardware characteristics from the particular post-processing algorithms paired with each sensor. The absence of error bars and direct measurements of the proposed degradation mechanisms further weakens the causal claims. The framework is a promising contribution, but the central claim requires additional control experiments before it is fully supported.","major_comments":[{"comment":"The comparison protocol is asymmetric in a load-bearing way. Tianmouc is evaluated with the authors' own reconstruction, interpolation, and optical-flow code (refs. [2], [27], [28]), including dynamically adjusted sampling intervals that align with rotation increments exceeding 1 degree, whereas the event cameras use fixed third-party tools (ARC*, EV-FlowNet, E2VID, TimeLens) with fixed integration windows of 1.5 and 15 degrees. The observed gap could therefore reflect algorithmic maturity or tuning rather than sensor hardware. The paper acknowledges in Section VI that event algorithms are advancing and that Tianmouc software is immature, but it does not provide any sensitivity analysis or crossover experiment. I request a matched-algorithm evaluation: for example, feeding AOP-SD data through an event-based corner or flow algorithm, or sweeping the event integration windows and using more recent event-based algorithms, to show that the performance difference survives when algorithm choice is controlled.","section":"Section V.B, Section IV.E, Tables I–III"},{"comment":"No error bars, trial counts, or statistical tests are reported anywhere in the quantitative evaluation. Several comparisons show small differences (for example, F1 scores of 0.579 vs. 0.720 at 100 rpm in Table III), and without variance estimates or repeated trials it is impossible to determine whether the reported differences are significant. For a paper whose contribution is a quantitative evaluation framework, reporting the number of repeated runs and confidence intervals or performing a significance test is essential.","section":"Figures 4–6 and Tables I–III"},{"comment":"The causal attribution of EVS degradation to 'pixel bandwidth variation' and 'event rate saturation' is inferred rather than directly demonstrated. The DVXplorer ROI experiment is suggestive, but the paper never reports measured event rates, dropped events, pixel latency, or bus utilization. Please provide direct measurements or controlled manipulations that isolate these mechanisms, or alternatively reframe the conclusions as consistent hypotheses rather than established causes.","section":"Section VI, Section IV.D"},{"comment":"The definition of ground truth for corner detection needs clarification. The text states that detected corners are matched against ground-truth corner positions 'obtained using the Shi-Tomasi algorithms on original images.' It is unclear what 'original images' refers to for each sensor, whether the same ground-truth set is used across sensors, and whether using the same detector family for both Tianmouc and ground truth biases the comparison. Please specify the ground-truth source precisely and, if possible, use fixed printed-pattern coordinates as ground truth.","section":"Section V.A"}],"minor_comments":[{"comment":"The edge-thickness metric is described as 'the peak width (from maximum intensity to near zero)' but the exact threshold and computation procedure are not specified; please provide a precise definition.","section":"Section IV.C"},{"comment":"The N.A. entries for DAVIS346 at 300 rpm and above are unexplained; a footnote should state why no values are reported.","section":"Tables I–III"},{"comment":"Several subfigures lack clear axis labels and legend identifiers, and the difference between panels (a), (b), (c), and (d) is not always obvious; please improve the figure captions and panel labels.","section":"Figures 4–6"},{"comment":"The phrase 'the approach described in [2]' is too vague for a quantitative comparison; please specify the optical-flow algorithm, its hyperparameters, and the exact dynamic-frame-interval rule used for Tianmouc.","section":"Section V.B"},{"comment":"The paper does not include a data-availability statement; for a benchmark-oriented contribution, releasing the recorded sensor data and evaluation code would substantially aid reproducibility.","section":"General"}],"recommendation":"major_revision","confidential_remarks":"The Tianmouc sensor, its post-processing code, and the evaluation all originate from the same group, while the event cameras are paired with fixed third-party algorithms. This asymmetry is the core technical concern and should be addressed with matched-algorithm or crossover experiments. The paper would also benefit from independent validation or third-party data. I see no evidence of misconduct, but the current manuscript is not yet suitable for publication without the additional controls."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Short version: this is the first quantitative head-to-head between event cameras and the Tianmouc primitive-based sensor under controlled high-speed rotation, and the setup is genuinely well thought-out: a turntable with calibrated speeds, two illumination levels, motion-compensated event warping for fair edge recovery, and multiple metrics (edge thickness, structural fidelity, BRISQUE, corner detection, optical flow). Credit where due: the trends are internally consistent across metrics — EVS drops sharply in cluttered high-speed conditions, Tianmouc stays flat — and the paper's discussion is honest about algorithmic immaturity on both sides.\n\nThe problem is that the central causal claim — that EVS degradation is due to pixel bandwidth variation and event rate saturation, and Tianmouc's robustness stems from its spatiotemporal gradient sampling — is not actually established. The comparison pairs each sensor with a different post-processing pipeline. Tianmouc gets the authors' own reconstruction, interpolation, and flow code, with dynamically adapted integration windows; event cameras get third-party fixed tools (ARC*, EV-FlowNet, E2VID, TimeLens) with fixed 1.5 and 15 degree windows. The 15 degree window at high speed can blur flow estimates all by itself. The paper's own Discussion admits this is a limitation and calls for 'more representative algorithms' as future work, but no sensitivity analysis, crossover experiment, or matched-algorithm baseline is provided. So the headline takeaway about sensor selection is not yet evidence.\n\nAlso: no error bars, no trial counts, no data release. The Tianmouc sensor, its code, and the evaluation come from the same lab, which makes the asymmetry harder to wave away. The reader's stress-test note captures this accurately; I agree with it.\n\nWho is this for? Researchers choosing sensors for high-speed robots, and chip developers working on event-based or primitive-based designs. They'll get a useful map of the problem space and a framework to build on, even if they should not take the hardware conclusions at face value. I would send it out to referees, but the verdict should be major revision: require matched or at least crossed algorithm pipelines, add trials and error bars, and commit to releasing data and code. With that, the comparison could be a real reference point.","headline":"Useful first cross-BVS benchmark, but the hardware-level conclusion is undercut by an algorithm confound and missing reproducibility data.","tokens_in":12586,"tokens_out":2210,"would_cite":true,"duration_ms":22208,"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":"Event cameras degrade in high-speed clutter; Tianmouc does not.","keywords":["brain-inspired vision sensors","event-based vision","Tianmouc","high-speed robotic perception","motion blur","corner detection","optical flow","sensor benchmarking"],"falsifier":"Re-run the turntable protocol on a cluttered QR pattern using an event camera's raw streams but with the most recent learned corner and optical-flow algorithms in place of ARC* and EV-FlowNet; if event-camera F1 and angular-velocity error no longer decline with speed, the central claim that EVS intrinsically degrades in high-speed clutter is falsified.","tokens_in":11602,"feed_emoji":"👁️","tokens_out":12551,"duration_ms":115412,"temperature":0.7,"pith_summary":"Brain-inspired vision sensors promise high temporal resolution for robots, but there has been no standard way to compare them fairly. This paper builds that standard—a calibrated turntable, printed ground-truth patterns, motion-compensated event calibration, and shared imaging and perception metrics—and uses it to compare an event camera, an RGB-hybrid event camera, and the primitive-based Tianmouc chip. Its central finding is that event-based sensors keep their accuracy in fast sparse scenes and in moderately fast complex scenes, but lose structural information in high-speed cluttered scenes because pixel-level bandwidth varies and event output saturates. Tianmouc, which samples both temporal and spatial differences synchronously at up to 1515 frames per second, keeps stable structural metrics, corner detection, and motion estimation across speeds and scene complexity. If the finding holds, choosing a vision sensor for high-speed robots becomes a scenario-dependent decision rather than a one-class verdict.","feed_headline":"Event cameras falter in fast clutter; Tianmouc stays sharp","feed_subtitle":"A turntable benchmark shows sensor choice for high-speed robots depends on scene clutter.","key_machinery":"The argument is carried by a side-by-side comparison of two data modalities. Event cameras are modeled by the event generation condition $\\Delta L(x_k,y_k,t_k)=p_k C$, so a pixel emits an asynchronous event only when log-intensity crosses a contrast threshold; this is the mechanism that, under load, saturates and loses structure. Tianmouc's AOP instead computes multi-bit temporal difference $TD(x,y,t)=I(x,y,t)-I(x,y,t-1)$ and spatial differences such as $SD_{\\mathrm{LeftDir}}(x,y,t)=I(x,y,t)-I(x-1,y-1,t)$ on a synchronous global shutter at up to 1515 fps, giving direct spatial-gradient output that does not wait for temporal contrast events. Around that comparison, the paper builds a unified evaluation platform—a turntable with adjustable rotation speed, printed patterns as ground truth, motion-compensated event warping using rotation speed as flow ground truth, and shared metrics (edge thickness, GM/TSS/VAR/GradVar, BRISQUE, corner F1, angular-velocity error)—so that differences in these metrics are attributed to the sensing scheme.","core_discovery":"The paper claims that, under a unified turntable-based protocol with motion-compensated calibration, event-based vision sensors (EVS) preserve data fidelity in high-speed, sparse scenes and in moderately fast, complex scenes, but degrade sharply in high-speed, cluttered scenes. The degradation is traced to pixel-level bandwidth variation under low light and to event-rate saturation when the on-chip bus is overloaded by dense features, and it shows up in edge thickness, structural-information metrics (GM, TSS, VAR, GradVar), BRISQUE reconstruction scores, and corner-detection F1. By contrast, Tianmouc's action-oriented pathway, which synchronously samples multi-bit temporal differences and spatial differences at up to 1515 fps with a global shutter, maintains stable structural metrics, corner detection, and angular-velocity estimation across sparse and complex scenes and across speeds. The paper presents this as evidence that sensor choice for high-speed robotics should depend on scene complexity and speed, and that Tianmouc's spatiotemporal-gradient sampling is a robust alternative where event cameras saturate.","pith_inferences":["A testable extension is to rerun the same turntable protocol with newer learned event-based corner detectors and optical-flow networks; if those close the gap with Tianmouc in the cluttered high-speed regime, part of the reported advantage belongs to software, not to the AOP sampling scheme.","The paper's own discussion implies that in low-light and ultra-high-speed conditions every tested sensor blurs because of pixel sensitivity; pushing the protocol into those regimes could narrow or invert Tianmouc's advantage.","Because the paper normalizes structural metrics to low-speed values, an absolute reference—say, a global-shutter frame camera on the same turntable—could turn these relative comparisons into a calibration-style standard for brain-inspired sensors.","A simpler design rule follows: any sensor that outputs spatial differences directly (gradients) should be less sensitive to edges aligned with the direction of motion than a purely temporal-contrast sensor; this can be isolated by rotating a pattern with edges parallel to motion and comparing TD-only and SD-only streams from the same chip."],"forward_implications":["For high-speed robots operating in sparse scenes, an event camera remains a low-bandwidth, low-power choice because its corner and flow accuracy holds when few features compete for event bandwidth.","In high-speed cluttered scenes, event-based pipelines should not be assumed robust: the paper measures sharp drops in structural metrics and corner-detection F1 as rotation speed rises.","Tianmouc's AOP-SD mode offers an alternative for those conditions, with F1 near 0.7 across 50–500 rpm and stable angular-velocity estimates at high speed.","When an event camera must be used at high speed, widening the temporal integration window can recover motion estimation, but at the cost of extra latency.","The evaluation protocol itself—turntable ground truth, motion-compensated calibration, and normalized structural metrics—can be reused to benchmark future brain-inspired sensors."],"supporting_citations":[{"why":"Supplies the event generation model and temporal-contrast sensing principle that define EVS in the comparison.","marker":"[1]"},{"why":"Defines Tianmouc's COP/AOP architecture and TD/SD equations, and provides the reconstruction/interpolation approach used for Tianmouc in the benchmarking.","marker":"[2]"},{"why":"Documents the loss of information when purely temporal sensors face edges aligned with motion, one of the three EVS limitations the paper invokes.","marker":"[7]"},{"why":"Establishes the event-rate saturation and output-bus overload mechanism the paper uses to explain EVS degradation in cluttered high-speed scenes.","marker":"[12]"},{"why":"Provides the contrast-maximization/warping formulation used to motion-compensate events before edge-quality measurement.","marker":"[23]"},{"why":"E2VID is the event-to-intensity reconstruction network whose output is scored with BRISQUE in the imaging-quality comparison.","marker":"[26]"},{"why":"Shi-Tomasi is the detector applied to Tianmouc's reconstructed intensities, producing both Tianmouc detections and ground-truth corners for F1 matching.","marker":"[31]"},{"why":"ARC* is the asynchronous corner detector whose detections are matched against ground truth to produce event-camera corner-detection F1.","marker":"[32]"},{"why":"EV-FlowNet supplies the neural-network optical-flow baseline for event-based motion estimation.","marker":"[33]"}],"fun_headline_variants":["Event sensors choke in fast clutter; Tianmouc endures","New benchmark: event cameras lose to Tianmouc in cluttered scenes","High-speed robot vision: Tianmouc beats event cameras in clutter","Event rate saturation: why Tianmouc wins in dense scenes","Sensor choice matters: Tianmouc robust where events saturate"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that the post-processing algorithms paired with each sensor are representative enough that performance differences reflect sensor hardware rather than algorithmic maturity—Tianmouc is processed with the authors' own reconstruction, interpolation, and optical-flow code, while event cameras are processed with fixed third-party tools.","fun_headline_variants_meta":{"raw":{"variants":["Event sensors choke in fast clutter; Tianmouc endures","New benchmark: event cameras lose to Tianmouc in cluttered scenes","High-speed robot vision: Tianmouc beats event cameras in clutter","Event rate saturation: why Tianmouc wins in dense scenes","Sensor choice matters: Tianmouc robust where events saturate"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000183,"raw_usage":{"total_tokens":1352,"prompt_tokens":1019,"completion_tokens":333,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":635,"completion_tokens_details":{"reasoning_tokens":244}},"tokens_in":635,"tokens_out":333,"duration_ms":3424,"temperature":1.0,"reasoning_tokens":244,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-16T05:57:43.516010+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Re-run the turntable protocol on a cluttered QR pattern using an event camera's raw streams but with the most recent learned corner and optical-flow algorithms in place of ARC* and EV-FlowNet; if event-camera F1 and angular-velocity error no longer decline with speed, the central claim that EVS intrinsically degrades in high-speed clutter is falsified.","supporting_citations":[{"cited_title":"A vision chip with complementary pathways for open-world sensing,","cited_arxiv_id":null,"evidence_quote":"Defines Tianmouc's COP/AOP architecture and TD/SD equations, and provides the reconstruction/interpolation approach used for Tianmouc in the benchmarking."},{"cited_title":"Microsaccade-inspired event camera for robotics,","cited_arxiv_id":null,"evidence_quote":"Documents the loss of information when purely temporal sensors face edges aligned with motion, one of the three EVS limitations the paper invokes."},{"cited_title":"High speed and high dynamic range video with an event camera,","cited_arxiv_id":null,"evidence_quote":"E2VID is the event-to-intensity reconstruction network whose output is scored with BRISQUE in the imaging-quality comparison."},{"cited_title":"Asynchronous corner detection and tracking for event cameras in real time,","cited_arxiv_id":null,"evidence_quote":"ARC* is the asynchronous corner detector whose detections are matched against ground truth to produce event-camera corner-detection F1."},{"cited_title":"Ev-flownet: Self-supervised optical flow estimation for event-based cameras,","cited_arxiv_id":null,"evidence_quote":"EV-FlowNet supplies the neural-network optical-flow baseline for event-based motion estimation."}],"review_version":1}