{"id":"8b1fea92-9708-41ec-a2ca-6dc9c70fb68e","arxiv_id":"2412.05053","paper_version":3,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":7.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":0,"one_line_summary":"EvTTC provides the first event-camera dataset with ground-truth TTC for high-relative-speed, emergency-braking driving scenarios, plus a small-scale testbed.","lead":"EvTTC is a new open dataset for training and testing time-to-collision (TTC) algorithms using event cameras, with synchronized RGB, LiDAR, and GPS/INS ground truth recorded in emergency braking scenarios. It is aimed at making automatic emergency braking systems react faster to sudden high-speed collision risks than frame-based cameras allow.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Ground-truth TTC accuracy rests on unmeasured synchronization and FAST-LIO2 depth drift; no independent validation is provided, so the central dataset claim is not yet demonstrated.","rationale":"The reader's weakest assumption identifies precisely the load-bearing condition: the utility of the dataset depends on the ground-truth TTC being trustworthy, and the paper never demonstrates that the timestamps of the LiDAR-derived depth and the GNSS/INS-derived velocity are aligned tightly enough for the high-speed emergency scenarios. I agree with that assessment. The manuscript invokes the sub-microsecond capability of PTP/gPTP but does not report a measured synchronization error for the actual sensor rig, and it relies on FAST-LIO2 odometry for depth accumulation without a drift check. The proposed consistency test would settle the question using only the released data: over a constant-velocity interval, the time derivative of the depth of the collision target must equal the negative of the reported relative velocity if the GT is unbiased. This is a standard, self-contained validation that directly targets the causal link between synchronization/depth errors and GT TTC bias. The paper remains a plausible and useful dataset contribution, and the absence of such validation is an addressable gap rather than a fatal flaw; hence the conditional verdict stands unchanged.","tokens_in":11944,"tokens_out":5328,"duration_ms":58196,"concrete_test":"On a constant-speed segment of a released sequence, extract the collision-target depth Z from the GT depth maps, compute its smoothed time derivative dZ/dt, and compare with the reported longitudinal relative velocity -Vrel from GNSS/INS. Fit a time shift delta to best align the two curves. If |delta| exceeds 2 ms, or the residual is above the combined LiDAR/GNSS error budget, then the GT TTC in Eq. (1) is biased by synchronization or depth-drift errors.","verdict_should_be":"UNCHANGED","load_bearing_attack":"EvTTC's central value is that Eq. (1) yields accurate ground-truth TTC. That accuracy requires Z (LiDAR-derived depth, accumulated with FAST-LIO2 poses) and Vrel (GNSS/INS velocity) to be measured at the same instant and without bias. The paper only states that PTP/gPTP 'can provide sub-microsecond synchronization' and that a microcontroller emits 20 Hz pulses; it reports no measured end-to-end timestamp error between LiDAR points, GNSS/INS samples, and camera exposures, nor any validation of FAST-LIO2 odometry drift on these sequences. In the emergency-braking sequences, Vrel reaches roughly 70 km/h and GT TTC values are near 1 s, so a few milliseconds of misalignment or a few centimeters of depth drift produce GT errors of the same order as the best benchmark eTTC values (e.g., CMax 2.56%). Because no independent check against a total station, high-speed camera, or known geometry is presented, the GT TTC could be silently biased without affecting any reported number. This does not invalidate the dataset, but it makes the 'accurate ground truth' part of the central claim conditional on an unverified assumption.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper introduces EvTTC, a multi-sensor dataset for time-to-collision estimation with event cameras, targeting high-relative-speed and emergency-braking driving scenarios. The sensor suite comprises two rigidly paired RGB/event camera rigs with 8-mm and 16-mm lenses, a Livox HAP LiDAR, and two GNSS/INS units. Ground-truth TTC is defined as TTC = Z/Vrel, where Z is the target depth from LiDAR point clouds (accumulated with FAST-LIO2 poses) and Vrel is the relative velocity from GNSS/INS. The dataset includes car-to-car and car-to-pedestrian sequences based on Euro NCAP AEB test protocols, plus 2D bounding boxes and depth maps. A small-scale linear-rail testbed with a 1:24 vehicle model is described as a low-cost complement. A benchmark evaluates six TTC estimation methods (including the authors' STRTTC) on selected sequences and the testbed.","tokens_in":12102,"tokens_out":3950,"duration_ms":42284,"significance":"If the ground-truth quality claims are substantiated, EvTTC would fill a clear gap: existing event-camera driving datasets (MVSEC, DSEC, ViViD++, M3ED) lack collision-focused, high-relative-speed scenarios with TTC ground truth. The open-sourced testbed and benchmark are useful community assets, and the use of independent sensors (LiDAR, GNSS/INS) rather than self-generated labels for TTC ground truth is a meaningful strength. The scenario design follows established Euro NCAP protocols, and the paper includes reproducible benchmark code and results. The main weakness is that the accuracy of the central ground-truth TTC is asserted rather than demonstrated; no measured synchronization error or odometry-drift validation is reported.","major_comments":[{"comment":"The ground-truth TTC is defined as TTC = Z/Vrel, which is the time to collision under a constant relative speed. However, the scenario parameters in Table III include a braking phase ΔT2 lasting up to 5.5 s, and the abstract and introduction emphasize 'emergency braking' with 'rapid decrease in vehicle speed.' During the braking phase Vrel is not constant, so the actual collision time under the recorded deceleration profile differs from Z/Vrel. The authors should clarify whether the ground-truth TTC is intentionally the instantaneous constant-velocity TTC (as the introduction's 'under their current speed' suggests) or the true collision time under continued braking, and, if the former, justify its use for AEB/FCW benchmarking or provide the deceleration profile so users can compute the latter. This is load-bearing because all benchmark eTTC errors are computed against this definition.","section":"Sec. IV-B, Eq. (1), Table III"},{"comment":"The paper claims sub-microsecond synchronization via PTP/gPTP and uses FAST-LIO2 poses to accumulate LiDAR point clouds for depth, but it provides no measured end-to-end synchronization error between the LiDAR points, the GNSS/INS samples, and the camera/event exposures, and no evaluation of FAST-LIO2 drift on these sequences. At the reported relative speeds (up to 70.4 km/h) and small ground-truth TTC values (near 1 s), a few milliseconds of temporal misalignment or a few centimeters of depth drift would produce TTC errors on the same order as the best benchmark eTTC values (e.g., 2.56% for CMax on CCRs1-low). The accuracy of the ground-truth TTC is therefore conditional on unverified assumptions. The authors should add a synchronization validation (e.g., a controlled LED flash or comparison of the 20 Hz trigger pulses with observed event timestamps) and an odometry drift check by comparing FAST-LIO2 poses with the RTK GNSS/INS trajectory over each sequence.","section":"Sec. III-B and Sec. IV-B"},{"comment":"The exact data flow from raw LiDAR and GNSS/INS measurements to the scalar Z and Vrel used in Eq. (1) is underspecified. The text says that the position of a stationary target is determined by measuring the distance using LiDAR, but the subsequent paragraph describes generating dense depth maps by accumulating velocity-compensated point clouds with FAST-LIO2 and applying Hidden Point Removal, then projecting into the camera coordinate frame. It is unclear whether the target-specific Z in Eq. (1) is taken from a single LiDAR return, the accumulated depth map at the target's 2D bounding box, or some other selection, and which timestamp is associated with Z and Vrel. A precise description of this pipeline (including the coordinate frame and the temporal alignment of Z and Vrel) is needed for the ground-truth generation to be reproducible and for users to assess its accuracy.","section":"Sec. IV-B, 'Pose and Depth'"}],"minor_comments":[{"comment":"The 'Detection Range [m]' entries for EvTTC ([160-295] and [99-197]) are not defined in Table II or the text; please clarify what these ranges represent (e.g., object detection range of the camera pair? LiDAR range?) and how they were measured.","section":"Table I"},{"comment":"The sentence 'hardware triggered synchronization is not witnessed' for ViViD++ should be rephrased, e.g., 'hardware-triggered synchronization is not reported.'","section":"Sec. II.3"},{"comment":"The description 'These pulses are used to simultaneously trigger the two RGB cameras and the two event cameras' is unclear because event cameras are asynchronous sensors; clarify whether the pulses are used to timestamp the event streams or to trigger a reset/external signal, and report the measured pulse-to-pulse jitter.","section":"Sec. III-B"},{"comment":"Please define the sign convention for Vrel: is it positive when the target is approaching, and is the TTC always positive? Also specify whether Z is the depth of the collision target in the 8-mm RGB camera frame, which appears to be the reference frame, or the event camera frame.","section":"Sec. IV-B, Eq. (1)"},{"comment":"The runtime for ETTCM is described as the product of per-event computation time and total event count; this should be stated directly in the table caption or in the text preceding the table for clarity.","section":"Sec. VI, Table V"},{"comment":"The benchmark includes the authors' own method STRTTC, which achieves the best or second-best results on most sequences. This is acceptable for a dataset paper, but the text should state whether the authors ran all methods or whether the results were independently reproduced, and note any potential bias in the methodology or parameter tuning.","section":"Sec. VI"}],"recommendation":"major_revision","confidential_remarks":"The paper makes a valuable contribution to the event-camera dataset literature, and the open-sourced testbed and benchmark are appreciated. The main obstacle to acceptance is the lack of direct validation of the ground-truth TTC accuracy, specifically the synchronization error and FAST-LIO2 odometry drift. These are fixable with additional experiments and clarifications. The inclusion of the authors' own STRTTC method in the benchmark is not a circularity problem because the ground truth comes from independent sensors, but I would encourage the authors to disclose parameter-tuning details and, if possible, have one baseline run by a third party. If the validation experiments are added and the ground-truth definition is clarified, the paper would be a solid contribution to RA-L."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"EvTTC is a genuinely useful new resource: the first event-camera dataset that combines TTC ground truth with emergency-braking, high-relative-speed scenarios, and an open-source small-scale testbed. The sensor rig is sensible—two event/RGB pairs, LiDAR, GNSS/INS—and the ground-truth TTC is computed from LiDAR depth and GNSS/INS velocities, which are independent of the evaluated algorithms, so the central dataset claim is not circular.\n\nThe main soft spot is that ground-truth accuracy is asserted, not demonstrated. The paper states that PTP/gPTP 'can provide sub-microsecond synchronization,' but reports no measured end-to-end timestamp error between LiDAR, GNSS/INS, and camera exposures. The depth comes through FAST-LIO2 odometry, and no drift validation is given on these sequences. In the 70 km/h cases, a few milliseconds of misalignment or a few centimeters of depth error would shift TTC by the same magnitude as the best benchmark errors. That makes the 'accurate ground truth' claim conditional on an unverified assumption. It is an addressable issue—add a sync validation experiment and an independent ground-truth check (e.g., against a total station or known geometry)—not a fatal flaw.\n\nMinor points: the benchmark table covers only a subset of the sequences, reports no error bars or number of runs, and includes the authors' own STRTTC method ranking best or second best. That's not disqualifying, but a fuller table with variance would strengthen it. Also, the constant-relative-speed TTC model needs a sentence clarifying how Z and Vrel are aligned during the braking phases.\n\nOverall, the paper fills a real gap and the data are external to the algorithms. I'd send it to review and, after the validation experiments, accept it. The event-vision and AEB communities would get value from the dataset and testbed.","headline":"Useful first event-camera TTC dataset under emergency braking, but ground-truth accuracy needs validation.","tokens_in":12715,"tokens_out":2895,"would_cite":true,"duration_ms":28382,"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":"This paper introduces EvTTC, the first event-camera dataset built for time-to-collision estimation in high-relative-speed, emergency-braking driving scenarios, with LiDAR/GNSS ground truth and an open small-scale testbed.","keywords":["event camera","time-to-collision estimation","autonomous driving dataset","forward collision warning","emergency braking","LiDAR-INS ground truth","event-based vision benchmark","small-scale testbed"],"falsifier":"Set up a repeating emergency-braking run with a stationary target at known distance and instrumented speed, and compare the dataset's published TTC values against an independent reference computed from a high-rate motion-capture system and mechanical encoder; also measure the actual trigger-to-event timestamp offset via an electrical pulse injected into all sensors. If the TTC deviation grows with relative speed in a way consistent with a fixed time offset, the synchronization assumption is disproven.","tokens_in":11697,"feed_emoji":"⚡","tokens_out":7606,"duration_ms":62950,"temperature":0.7,"pith_summary":"This paper is building a benchmark: it claims that before EvTTC no public multi-sensor dataset paired event cameras with emergency-braking scenarios for time-to-collision estimation, and that this gap blocks progress on forward-collision-warning systems that need microsecond-level response. It supplies synchronized RGB, event, LiDAR, and GNSS/INS recordings of car-to-car and car-to-pedestrian approaches, with ground-truth TTC computed from the ratio of depth to relative speed. It also provides an open-source small-scale testbed for cheap, controlled collision experiments. If the dataset works as claimed, it gives the event-camera community a common evaluation ground for TTC methods in the extreme cases where frame-based cameras are weakest.","feed_headline":"New dataset benchmarks event-camera time-to-collision in emergencies","feed_subtitle":"LiDAR and GNSS ground truth plus an open testbed cover high-relative-speed braking scenes.","key_machinery":"The load-bearing object is the EvTTC dataset itself: a hardware-synchronized sensor suite with two RGB-event camera pairs (8-mm and 16-mm lenses), a Livox LiDAR, and dual GNSS/INS units, covering Euro NCAP AEB car-to-car and car-to-pedestrian scenarios. Event cameras are the sensors that asynchronously report per-pixel brightness changes at microsecond resolution, which is what the paper argues gives them an advantage over frame-based cameras in sudden braking cases. The identity that defines ground truth is TTC = Z / Vrel, where Z is the target depth in the camera frame and Vrel is the relative speed along the optical axis; Z comes from LiDAR-inertial odometry (FAST-LIO2) with hidden-point removal, and Vrel comes from GNSS/INS at 100 Hz. The synchronization scheme, PTP/gPTP plus micro-controller trigger pulses at 20 Hz, is what ties the sensors' timestamps together. The small-scale testbed adds a linear rail, motor encoder, and beam-splitter optical system as a second, repeatable source of ground-truth TTC.","core_discovery":"On its own terms, the paper's contribution is a new dataset and benchmark rather than a new TTC algorithm. The central claim is that EvTTC fills the missing piece in TTC research: a public, multi-sensor event-camera dataset whose sequences are designed around the high-relative-speed emergency-braking scenarios that frame-based ADAS cameras handle poorly. Ground-truth TTC is defined directly by TTC = Z/Vrel, with depth from LiDAR and relative velocity from GNSS/INS, and every sequence carries depth, pose, and 2D bounding-box annotations. The dataset also comes with a small-scale testbed that generates quasi-real collision data at controlled speeds. The reported benchmark shows existing event-based TTC methods achieving errors mostly in the single-digit to tens-of-percent range on these sequences.","pith_inferences":["By releasing per-event timestamps alongside ground-truth TTC, the dataset could be used to compute a latency-aware metric such as the latest sensor time at which a correct TTC estimate is still available, a step the paper does not take.","The high-speed cases make the ground truth sensitive to synchronization offset, so a natural robustness check is to recompute TTC under deliberately shifted timestamps and report error growth; this would show how much of the benchmark results depends on the sync claim.","The 1:24 scale testbed and motor-encoder ground truth could be used for controlled studies of event-camera latency in TTC estimation, since the linear rail provides a repeatable trajectory with known ground truth.","The benchmark's frame-based FoE baseline performs well in several sequences; an extension the authors do not explore is measuring how much worse it becomes as frame rate is lowered, which would quantify the latency advantage of event cameras."],"forward_implications":["Event-based TTC algorithms can now be evaluated under emergency-braking conditions with ground-truth TTC values, not just on normal-driving scenes or synthetic data.","The benchmark gives a direct numerical comparison of six TTC estimators, so a method's accuracy and runtime on high-relative-speed sequences become a public, repeatable result.","The small-scale testbed lowers the cost of testing and augmenting TTC algorithms before full-scale vehicle experiments, while still providing precise motor-encoder ground truth.","Because the suite includes RGB, event, LiDAR, and GNSS/INS streams, the same scenarios support future multi-sensor fusion work for forward collision warning.","The scenario design follows Euro NCAP AEB test protocols, so results on EvTTC relate directly to the kinds of tests used in vehicle safety assessment."],"supporting_citations":[{"why":"The prior TTC-specific dataset, used to show existing TTC data lack high-speed emergency braking.","marker":"TSTTC [4]"},{"why":"The comparison baseline event dataset that lacks TTC ground truth and emergency braking scenarios.","marker":"MVSEC [5]"},{"why":"A leading stereo event dataset that provides depth and RTK GPS but no TTC task.","marker":"DSEC [6]"},{"why":"The high-resolution multi-sensor event dataset used to motivate the need for a TTC-focused collection.","marker":"M3ED [8]"},{"why":"Euro NCAP AEB test protocol defines the Car-to-Car and Car-to-Pedestrian scenarios that structure EvTTC's sequences.","marker":"[38]"},{"why":"The LiDAR-inertial odometry that generates ground-truth poses and velocity-compensated point clouds used for depth and TTC.","marker":"FAST-LIO2 [39]"},{"why":"The top-performing event-aided TTC estimator benchmarked on the dataset.","marker":"STRTTC [32]"},{"why":"The contrast-maximization baseline that the dataset's benchmark evaluates for TTC accuracy.","marker":"CMax [28]"}],"fun_headline_variants":["Event-camera TTC dataset targets high-relative-speed emergencies","New open dataset + testbed for event-camera TTC","First event-camera TTC dataset for sudden speed changes","Benchmarking event-camera TTC with LiDAR and GNSS ground truth"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that the sub-microsecond synchronization and LiDAR odometry align all measurements closely enough that depth Z and relative speed Vrel are measured at the same instant, with any misalignment or drift being negligible in the high-relative-speed cases where small time offsets produce large TTC errors.","fun_headline_variants_meta":{"raw":{"variants":["Event-camera TTC dataset targets high-relative-speed emergencies","New open dataset + testbed for event-camera TTC","First event-camera TTC dataset for sudden speed changes","Benchmarking event-camera TTC with LiDAR and GNSS ground truth"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000929,"raw_usage":{"total_tokens":4003,"prompt_tokens":995,"completion_tokens":3008,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":611,"completion_tokens_details":{"reasoning_tokens":2934}},"tokens_in":611,"tokens_out":3008,"duration_ms":193803,"temperature":1.0,"reasoning_tokens":2934,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-11T20:56:46.848306+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Set up a repeating emergency-braking run with a stationary target at known distance and instrumented speed, and compare the dataset's published TTC values against an independent reference computed from a high-rate motion-capture system and mechanical encoder; also measure the actual trigger-to-event timestamp offset via an electrical pulse injected into all sensors. If the TTC deviation grows with relative speed in a way consistent with a fixed time offset, the synchronization assumption is disproven.","supporting_citations":[{"cited_title":"Euro ncap aeb c2c test protocol - v4.3.1,","cited_arxiv_id":null,"evidence_quote":"Euro NCAP AEB test protocol defines the Car-to-Car and Car-to-Pedestrian scenarios that structure EvTTC's sequences."}],"review_version":1}