{"id":"fe54c061-7469-4f72-a63a-68985525b240","arxiv_id":"2505.05183","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":1,"one_line_summary":"Activated emergency vehicle lighting causes flare that makes object detector confidence fluctuate below detection thresholds in dark conditions, and a two-branch denoiser plus fine-tuned detector framework partially restores confidence at 30-50 FPS.","lead":"Flashing emergency vehicle lights can make car-detection AI confidence scores drop and jump, sometimes making the AI miss the emergency vehicle and nearby cars. The authors show this across seven ADAS camera systems and four object detectors, and they propose a real-time software mitigation called Caracetamol.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Caracetamol's headline gains may be inflated because Section 4.2.2 uses the same YouTube dataset for CycleGAN training and for the Section 4.5.1 evaluation, with no reported train/test split.","rationale":"The reader's weakest assumption was transferability to Tesla's proprietary detectors, which is explicitly disclaimed by the authors and does not undercut the empirically demonstrated phenomenon on public COCO detectors. The data-leakage issue, by contrast, directly threatens the quantitative headline of the mitigation, which is part of the central claim. The paper's own text raises the issue: the same YouTube corpus is used for CycleGAN training and for evaluation, and the limitations section acknowledges limited dataset availability; no split is documented. A concrete overlap check can settle it. I give credit for the PaniCar phenomenon evidence: multiple detectors, seven ADASs, FFT matching the photodiode frequency, and tracker results are independent and credible. The Caracetamol architecture also has a second gap, namely that the combiner described in Section 4.4 is never evaluated as part of the reported numbers, so the numbers describe a partial pipeline. I focus on the data-leakage concern because it is the most likely to change the magnitude of the claimed effect and is directly checkable. The reader's CONDITIONAL verdict remains appropriate, so I recommend no change.","tokens_in":25526,"tokens_out":6249,"duration_ms":63363,"concrete_test":"List the exact YouTube frame/video identifiers used as CycleGAN Group B (Section 4.2.2) and the subset used as the YouTube test set in Section 4.5.1; compute their overlap. If overlap is nonzero, re-run the 'Denoiser + YOLO MFA' and 'Denoiser + FRCNN MFA' rows of Table 4 on a held-out partition (e.g., 20% of the 243 videos) that was never used in CycleGAN training, and report average, minimum, and range. If the deltas shrink materially, the Caracetamol claims must be revised; if overlap is zero or the deltas persist, the concern is resolved.","verdict_should_be":"UNCHANGED","load_bearing_attack":"Section 4.2.2 states that the YouTube dataset 'was used to both evaluate Caracetamol and facilitate the training of a CycleGAN model,' and Section 4.5.1 reports 'We fed the YouTube data test to each model.' No split is described between the roughly 1,000 YouTube images used as CycleGAN Group B and the YouTube 'test' set. The denoiser is the main driver of the mitigation gain: fine-tuning alone moves YOLOv3 from 0.50 to 0.54, while adding the denoiser raises it to 0.71 (Table 4). If the denoiser was trained on the same videos on which it is evaluated, the reported +0.20 average-confidence, +0.33 lower-bound, and 0.33 range-reduction figures can reflect memorization rather than generalization. This does not impugn the PaniCar phenomenon itself, which is supported by multiple detectors, ADASs, and light patterns. It does mean the paper's central mitigation claim is not yet supported by a clean evaluation.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper identifies and characterizes PaniCar, a phenomenon in which activated emergency vehicle lighting causes the confidence scores of common object detectors to fluctuate widely over time, sometimes dropping below typical detection thresholds. The authors demonstrate this across four open-source COCO-pretrained detectors (YOLOv9/YOLOv3, SSD, RetinaNet, Faster R-CNN) applied to footage from seven commercial ADAS cameras and a smartphone, across 14 emergency-vehicle light patterns, and show via FFT analysis that the confidence fluctuation frequency matches the photodiode-measured modulation frequency of the lights. They further show that object trackers do not compensate for the detection loss, and that the effect is stronger in darkness. To mitigate PaniCar, the paper proposes Caracetamol, a pipeline consisting of a CycleGAN-based denoiser, a fine-tuned copy of the original detector, and a combiner, reporting large confidence improvements at 30-50 FPS, while also evaluating four state-of-the-art flare removal methods and finding them inadequate for real-time use. The authors explicitly disclaim access to Tesla's proprietary detectors and frame their analysis as based on commonly used detectors.","tokens_in":25623,"tokens_out":4868,"duration_ms":47216,"significance":"If it holds, the PaniCar phenomenon is an important empirical result for vision-based ADAS safety: it connects a known optical artifact (lens flare from dynamic emergency lighting) to a measurable, patterned degradation of object detector confidence, with plausible safety implications. The evidence for the phenomenon itself is strong and multi-faceted: it is reproduced across multiple detectors, multiple ADAS cameras, real recordings, and an external photodiode measurement, and the pattern-dependent detection-loss analysis shows a concrete mechanism. The dataset release and the negative evaluation of SOTA flare removal methods are also useful contributions. The Caracetamol mitigation, while promising, is undermined by a likely data-leakage problem in its evaluation and by inconsistencies between the paper's headline quantitative claims and the reported table; these issues must be resolved before the mitigation results can be accepted.","major_comments":[{"comment":"The Caracetamol evaluation appears to suffer from data leakage between CycleGAN training and evaluation. Section 4.2.2 states that the YouTube dataset 'was used to both evaluate Caracetamol and facilitate the training of a CycleGAN model,' and Section 4.5.1 reports 'We fed the YouTube data test to each model.' No train/test split is described between the approximately 1,000 YouTube images used as CycleGAN Group B and the YouTube 'test' set. This is load-bearing because the denoiser is the main driver of the reported improvement: in Table 4, fine-tuning alone moves YOLOv3 from 0.50 to 0.54, while adding the denoiser raises it to 0.71. If the denoiser was trained on the same videos on which it is evaluated, the reported gains (+0.20 average confidence, +0.33 lower bound, 0.33 range reduction) may reflect memorization rather than generalization. Please provide an explicit split (e.g., disjoint sets of YouTube videos for CycleGAN training and evaluation), re-run the Table 4 evaluation on the held-out set, and update all affected abstract and Section 7 numbers.","section":"4.2.2, 4.5.1, Table 4"},{"comment":"The paper's headline quantitative claims do not match the numbers in Table 4. For YOLOv3, the average confidence improves from 0.50 (YOLOCOCO) to 0.71 (Denoiser+YOLOMFA), which is +0.21, not +0.20; the absolute range reduces from 0.69 to 0.23, a reduction of 0.46, not 0.33. For Faster R-CNN, the average confidence improves from 0.63 to 0.81 (+0.18) and the minimum confidence from 0.21 to 0.45 (+0.24), not +0.33. The claim of 'reduces the fluctuation range by 0.33' is also not supported by the SSD row, where the range increases from 0.65 to 0.70. Please specify exactly which model/detector combinations support each claimed improvement, or revise the abstract and Section 7 to match the reported data.","section":"Abstract, Section 7, Table 4"},{"comment":"The paper motivates the work with the 16 documented Tesla crashes and states that PaniCar 'can cause autonomous vehicles to fail to detect objects near emergency vehicles.' However, all experiments use open-source COCO-pretrained detectors applied to footage from ADAS cameras, not the proprietary detectors actually deployed in Tesla or the other seven ADASs. The authors' own disclaimer in Sections 1 and 5 acknowledges that there may be a discrepancy, but the abstract and discussion do not carry that caveat. This is a load-bearing scope issue: the paper's safety-critical conclusion is about real ADAS perception, while the evidence is about a set of generic detectors on recorded footage. Please either (a) provide evidence that the tested detectors behave like production ADAS detectors (e.g., test on a production-like detection stack or on the ADAS's own detection outputs), or (b) explicitly reframe the conclusions as applying to common open-source detectors on ADAS camera footage and temper the language connecting the results to the documented Tesla crashes.","section":"Section 1, Section 2, Section 5"}],"minor_comments":[{"comment":"The dataset availability link is left as '??' and should be filled in.","section":"4.2.2"},{"comment":"There are typos, including 'lightning' for 'lighting' and 'Telsa' for 'Tesla' in the abstract; please proofread the final version.","section":"Abstract, throughout"},{"comment":"Table 5 uses green/red color coding to indicate whether 30-60 FPS is met; this may not be legible in black-and-white printing. Add explicit textual markers (e.g., 'yes'/'no') to each row.","section":"Table 5"},{"comment":"The 'Robustness against Adversaries' section is a brief argument rather than an experiment; consider renaming it to 'Discussion' or adding an actual adversarial evaluation if this claim is to be retained.","section":"4.5.3"},{"comment":"The metric 'Absolute Range' is used in Table 4 but not precisely defined in Section 4.2.1; please state that it is the maximum minus minimum confidence over the evaluated frames.","section":"Section 4.2.1, Table 4"},{"comment":"The caption repeats 'Tesla 2023 Tesla model 3'; please correct the wording.","section":"Figure 22 caption"}],"recommendation":"major_revision","confidential_remarks":"To the editor: The PaniCar phenomenon itself is well supported empirically and would be a valuable contribution. The main risk to the paper is the Caracetamol evaluation: if the reported train/test split is indeed absent, the mitigation results are not reliable and must be re-derived on a held-out set. The paper's framing around the Tesla crashes should also be made more cautious in the abstract and conclusion. I recommend requiring a clean evaluation of Caracetamol before acceptance; the phenomenon sections and the flare-removal comparison can likely stand with minor revisions."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The core finding here is real and worth knowing: flashing emergency vehicle lights make off-the-shelf object detectors' confidence scores oscillate at the same frequency as the light pattern, sometimes dipping below detection thresholds. That is a new empirical result, and the paper supports it with actual ADAS footage, multiple detectors, multiple ADASs, and a photodiode confirmation of the 1.3 Hz coupling. The evidence for the PaniCar phenomenon itself is solid, and the authors are appropriately careful about not overclaiming the link to the 16 Tesla crashes — they explicitly disclaim access to Tesla's detectors. Good on them for that, and for the responsible disclosure. The soft spot is the mitigation evaluation. Section 4.2.2 says the YouTube dataset was used both to train the CycleGAN denoiser and to evaluate Caracetamol, and no split is described. Section 4.5.1 then says 'we fed the YouTube data test to each model.' If the denoiser trained on the same videos it is evaluated on, the reported +0.20 average-confidence, +0.33 lower-bound, and 0.33 range-reduction figures could largely be memorization. That is load-bearing, because the denoiser is the main driver of the improvement in Table 4. The fine-tuning alone moves YOLOv3 from 0.50 to 0.54; adding the denoiser jumps to 0.71. This does not impugn the phenomenon, but it means the central mitigation claim is not yet cleanly supported. The abstract's aggregate numbers averaged across YOLOv3 and Faster R-CNN are also a bit loose — the reader has to dig into Table 4 to see the per-model deltas. Other issues are more minor: the dataset link is a placeholder, no code release, the mitigation experiments look like single-video runs without error bars, and the runtime numbers are on desktop GPUs, not embedded ADAS hardware. All addressable. This paper deserves a serious referee. The phenomenon is important enough, and the empirical demonstration is strong enough, that the Caracetamol evaluation problems should be fixed in review rather than used to reject outright. I would send it out, and I would tell the authors to provide a clean train/test split for the denoiser, release the data and code, and report per-model results with confidence intervals. If they can do that, this becomes a solid paper.","headline":"A genuinely new empirical finding about flashing emergency lights causing confidence fluctuations in object detectors, but the mitigation's headline numbers rest on a possible train/test overlap that needs fixing.","tokens_in":685,"tokens_out":1775,"would_cite":true,"duration_ms":34201,"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":"When emergency vehicle lights flash at night, object detectors' confidence scores for a nearby car swing wildly, sometimes below detection thresholds, and a software pipeline can largely stabilize them.","keywords":["PaniCar","object detection confidence","emergency vehicle lighting","lens flare","ADAS perception","autonomous driving safety","Caracetamol","detection threshold"],"falsifier":"Record the same nighttime scene with a production Level 2 ADAS's own vision stack and measure its reported confidence for a stationary vehicle in the path while emergency lights flash; if the confidence never dips below the deployment threshold, or the dips do not track the light pattern's frequency, then PaniCar as described does not transfer to the deployed system.","tokens_in":25236,"feed_emoji":"🚨","tokens_out":7395,"duration_ms":67368,"temperature":0.7,"pith_summary":"This paper argues that the flashing lights of parked emergency vehicles at night cause a previously unidentified failure in the vision systems that power advanced driving assistance systems (ADASs). The effect, which the authors call PaniCar, is a wide fluctuation in an object detector's confidence that a car is present, with the score periodically falling below a reasonable detection threshold. The paper documents the effect across seven commercial ADAS cameras, four common object detectors, and fourteen light patterns, and shows that the fluctuation tracks the frequency of the light pattern. It then proposes Caracetamol, a software framework that pairs a flare-removing denoiser with a fine-tuned copy of the detector, and reports that it raises average car-detection confidence by about 0.20, lifts the lower confidence bound by about 0.33, and shrinks the fluctuation range by about 0.33 on YOLOv3 and Faster R-CNN at 30–50 FPS. If correct, the finding gives a technical explanation for why autopilot-equipped cars repeatedly crashed into emergency vehicles, and offers a path to fix it.","feed_headline":"Flashing emergency lights make car AI miss vehicles at night","feed_subtitle":"PaniCar: flashing lights drop detector confidence below threshold; a software fix restores it in real time.","key_machinery":"The central object is PaniCar itself: a time-varying confidence signal produced by an object detector when flare from flashing emergency lights changes the tonal distribution of the car in successive frames, with the fluctuation's dominant frequency matching the light pattern's frequency (a 1.3 Hz peak in the paper's recordings, confirmed by photodiode measurements). The central mechanism carrying the mitigation is the Caracetamol framework, a four-part perception pipeline: a CycleGAN-based denoiser, i.e., a generative network trained to translate flashed frames back to unflashed frames, a fine-tuned replica of the ADAS's object detector running on the denoised frame, the original detector running on the original frame to preserve its detections, and a combiner that aggregates the two. The replication-plus-denoiser path supplies the confidences that the original detector loses, while the original path guarantees no regression on normal scenes.","core_discovery":"The paper's central claim is that activating emergency vehicle lighting creates a phenomenon, PaniCar, in which an object detector's confidence score for a detected object fluctuates within a wide range, dipping below a reasonable detection threshold in some frames. The underlying mechanism is the flare from the lights: the added light changes the tonal distribution of the car in the captured frame over time, and the pattern and frequency of the flashing lights dictate when confidence rises and falls. The paper shows that the fluctuation appears consistently across four object detectors (YOLO, SSD, RetinaNet, Faster R-CNN) and footage from seven ADAS cameras plus a smartphone, that object trackers layered on top do not compensate for the detection loss, and that the effect is strongest in darkness. It further finds that existing flare-removal methods either do not restore detector confidence or run too slowly for real-time driving, and proposes Caracetamol, a pipeline that combines a CycleGAN-based denoiser, a fine-tuned replica of the original detector, the original detector on the unprocessed frame, and a combiner layer. On YOLOv3 and Faster R-CNN, Caracetamol raises average car-detection confidence by 0.20, raises the lower confidence bound by 0.33, and reduces the fluctuation range by 0.33, at 30–50 FPS.","pith_inferences":["Because the confidence dips are periodic and phase-locked to the light pattern, a temporal filter that integrates confidence over one or two flash cycles could recover detection without a learned denoiser; the paper does not test this.","The spectral signature (around 1.3 Hz in the reported recordings) implies that the failure depends on the light pattern's frequency content, so patterns with faster or irregular flashing may produce different dip durations and severities than the ones measured here.","The transferability question cuts the other way too: production detectors with temporal smoothing or radar/camera fusion might already be partially robust, which would mean PaniCar is mostly a hazard for vision-only, per-frame systems rather than all ADASs.","If Caracetamol's gains hold on production hardware, the same denoise-then-finetune structure could generalize to other glare sources (oncoming headlights, sun glare) since the mechanism is a tonal shift that degrades confidence, not emergency lights specifically."],"forward_implications":["Vision-only ADAS and autopilot systems may intermittently fail to detect vehicles sitting near active emergency lights at night, even when the same vehicle is detected with high confidence a few frames earlier or later.","Object trackers do not repair the failure, so countermeasures must act at the detector level, not the tracking layer.","Existing flare-removal networks, which the paper evaluates, are too slow and too weak at restoring detector confidence to be used as-is in a real-time driving loop.","A detector-plus-denoiser pipeline such as Caracetamol can run within real-time budgets (31–50 FPS on the tested GPUs) while raising average confidence and shrinking the confidence range.","The same phenomenon can be weaponized: a person who places or triggers emergency-style flashing lights can cause an ADAS to miss a vehicle or obstacle, turning a safety issue into a deliberate attack surface."],"supporting_citations":[{"why":"Grounds the 16 documented crashes and the NHTSA investigation that motivate the study.","marker":"[1]"},{"why":"Supplies a state-of-the-art flare-removal baseline that the paper shows does not restore detector confidence in real time.","marker":"[7]"},{"why":"Supplies a second flare-removal baseline whose latency and accuracy are found insufficient for driving constraints.","marker":"[8]"},{"why":"Supplies the fastest flare-removal baseline tested, yet it still leaves average detector confidence below 0.6.","marker":"[9]"},{"why":"One of the four object detectors used to demonstrate the PaniCar confidence fluctuation.","marker":"[24]"},{"why":"The two-stage detector used throughout the analysis and as a base architecture for Caracetamol evaluation.","marker":"[25]"},{"why":"An object tracker implementation used to show that tracking does not compensate for detector-level detection loss.","marker":"[31]"},{"why":"Supplies the base driving dataset on which the augmented nighttime training images for Caracetamol are built.","marker":"[35]"},{"why":"Supplies the generative architecture used for both the Caracetamol denoiser and the synthetic emergency-light augmentation.","marker":"[37]"}],"fun_headline_variants":["Emergency vehicle lights make car AI miss vehicles, fix restores detection","PaniCar: Flashing emergency lights crash object detector confidence","Caracetamol rescues car detection from emergency light flare in real time","Flashing lights cause ADAS to miss cars; Caracetamol fixes it at 30 FPS"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The claim rests on assuming that the open-source, publicly trained object detectors used in the lab behave like the proprietary detectors inside the commercial ADASs — especially Tesla's — whose crash footage motivated the study, so the confidence dips observed here would actually occur in the deployed vehicles.","fun_headline_variants_meta":{"raw":{"variants":["Emergency vehicle lights make car AI miss vehicles, fix restores detection","PaniCar: Flashing emergency lights crash object detector confidence","Caracetamol rescues car detection from emergency light flare in real time","Flashing lights cause ADAS to miss cars; Caracetamol fixes it at 30 FPS"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000454,"raw_usage":{"total_tokens":2392,"prompt_tokens":1167,"completion_tokens":1225,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":783,"completion_tokens_details":{"reasoning_tokens":1143}},"tokens_in":783,"tokens_out":1225,"duration_ms":9276,"temperature":1.0,"reasoning_tokens":1143,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-15T23:10:57.602421+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Record the same nighttime scene with a production Level 2 ADAS's own vision stack and measure its reported confidence for a stationary vehicle in the path while emergency lights flash; if the confidence never dips below the deployment threshold, or the dips do not track the light pattern's frequency, then PaniCar as described does not transfer to the deployed system.","supporting_citations":[{"cited_title":"Odi resume,","cited_arxiv_id":null,"evidence_quote":"Grounds the 16 documented crashes and the NHTSA investigation that motivate the study."},{"cited_title":"Flare7k: A phe- nomenological nighttime flare removal dataset,","cited_arxiv_id":null,"evidence_quote":"Supplies a state-of-the-art flare-removal baseline that the paper shows does not restore detector confidence in real time."},{"cited_title":"Im- proving lens flare removal with general-purpose pipeline and multiple light sources recovery,","cited_arxiv_id":null,"evidence_quote":"Supplies a second flare-removal baseline whose latency and accuracy are found insufficient for driving constraints."},{"cited_title":"How to train neural networks for flare removal,","cited_arxiv_id":null,"evidence_quote":"Supplies the fastest flare-removal baseline tested, yet it still leaves average detector confidence below 0.6."},{"cited_title":"YOLOv9: Learning what you want to learn using programmable gradient information,","cited_arxiv_id":null,"evidence_quote":"One of the four object detectors used to demonstrate the PaniCar confidence fluctuation."},{"cited_title":"Unpaired image- to-image translation using cycle-consistent adversarial networks,","cited_arxiv_id":null,"evidence_quote":"Supplies the generative architecture used for both the Caracetamol denoiser and the synthetic emergency-light augmentation."}],"review_version":1}