{"id":"cd9af38e-631a-40e7-ba51-c50bc8841b07","arxiv_id":"2508.04564","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":2.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":0,"one_line_summary":"A survey of event camera-based drone detection that maps methods by data representation and covers tracking, forecasting, and propeller signature analysis.","lead":"This paper reviews how event cameras, which record brightness changes asynchronously, are being used to detect, track, and identify drones. It organizes current methods, datasets, and open problems, making a case for event-based vision as the backbone of future counter-UAV systems.","discovery_kind":"review","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Survey's central claim overreaches its own caveat: §1/Abstract promise robust performance in extreme lighting, but §3 concedes low light and rain/snow are 'challenging for event cameras too' (cited [9,38]).","rationale":"The reader's weakest assumption concerns the unverified transfer of event-camera properties to real drone-detection scenarios. This stress-test finds a more concrete, internal version of the same problem: the paper's own Section 3 states that low light and rain/snow are 'challenging for event cameras too,' directly contradicting the Abstract's unqualified 'consistent detection in extreme lighting' and Section 1's 'robust performance in extreme lighting conditions.' Since the central recommendation rests on these capabilities, the survey should either present stratified evidence from the datasets it reviews or qualify the claim to the conditions under which event cameras are known to perform well. The reader's conditional verdict already calls for qualification of overclaims, so my read does not change that verdict: conditional acceptance remains appropriate, with the specific qualification that extreme-lighting robustness is not established by the surveyed evidence and is internally contested. I do not see a need to reject or elevate the verdict; the paper is a survey and its core message can be repaired by aligning the Abstract/Conclusions with the limitations acknowledged in Section 3.","tokens_in":14892,"tokens_out":3772,"duration_ms":45189,"concrete_test":"Using the FRED dataset [38], compute the detection AP of the event-only baseline separately for the clear/backlight, rain, and low-light sequence subsets (as defined by the dataset's recording conditions). If AP in the rain or low-light subset is significantly lower than in the clear subset, the Abstract's 'consistent detection in extreme lighting' and the Conclusion's 'robust alternative' are not supported by the survey's own benchmark; the central claim should be qualified to conditions where event sensors are known to operate well. As a secondary check, read the two cited sources [9, 38] and extract their quantitative statements about rain/low-light degradation; if neither reports such measurements, the Sec. 3 caveat is unsupported too.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim—that event cameras are 'a powerful foundation' for counter-UAV systems because of microsecond temporal resolution, >120 dB dynamic range, and sparse output—is weakened by an internal contradiction. Section 1 asserts that a 'very high dynamic range of over 120 dB' enables 'robust performance in extreme lighting conditions,' and the Abstract promises 'consistent detection in extreme lighting.' Yet Section 3, 'Challenges of Drone Detection,' explicitly lists 'extreme scenarios' as including 'low light environments, and harsh meteorological conditions (eg. rain, snow) which are challenging for event cameras too [9, 38].' Low light is a lighting condition; if event cameras degrade there, the unqualified claim of consistent detection in extreme lighting is not supported by the survey's own cited evidence. HDR handles high intra-scene dynamic range (bright sky vs. dark drone), but not low absolute photon flux, where event thresholds and sensor noise dominate. The survey cites no quantitative comparison of event-based drone detection under these degradation conditions; the FRED dataset [38] is described as including rain and low light, but no stratified performance is reported. Thus the load-bearing assumption—sensor properties translate into reliable detection in the operational envelope—is not merely unproven; the paper itself supplies a countervailing caveat.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"This paper surveys event-based vision for drone detection and related counter-UAV tasks. It reviews sensor principles, event representations (event frames, point clouds, voxel grids, time-retaining frames, and end-to-end spiking pipelines), methods for detection, tracking, forecasting, and propeller-blade analysis, and it catalogs public and synthetic datasets in Table 1. The central claim is that event cameras' microsecond temporal resolution, >120 dB dynamic range, and sparse output make them a powerful foundation for reliable, low-latency, and efficient counter-UAV systems. The paper is primarily qualitative; it presents no new experiments or quantitative comparisons, but it offers a useful taxonomy and an organized overview of the literature.","tokens_in":15218,"tokens_out":7041,"duration_ms":84473,"significance":"If its claims are accepted, the survey consolidates a young and fast-moving field and could guide system-level design choices. Its categorization of event representations and tasks, the dataset comparison in Table 1, and the discussion of propeller-signature analysis are useful and broadly accurate. The survey also contains relevant caveats, such as the admission that low-light and rain/snow are challenging for event cameras (§3) and that frame-based accumulation can blur fast motion (§4). However, these caveats stand in tension with the unqualified claims in the Abstract and Introduction. Because the paper's headline advantages are asserted rather than demonstrated, and because the body itself weakens those assertions, the survey currently overstates its case. A revised version that separates sensor-level capabilities from system-level evidence and that verifies or softens exclusivity claims would significantly improve the paper.","major_comments":[{"comment":"The Abstract states that event cameras 'enable consistent detection in extreme lighting,' and §1 says the >120 dB dynamic range yields 'robust performance in extreme lighting conditions.' In direct tension, §3 lists 'low light environments, and harsh meteorological conditions (eg. rain, snow) which are challenging for event cameras too [9,38]'. Low light is a lighting condition, and rain/snow are precisely the extreme scenarios that a counter-UAV system must handle. Since this caveat concerns the exact capability used to motivate the survey, the headline claim should be narrowed (e.g., to high-contrast/backlit scenes) or supported by stratified results from datasets such as FRED [38], which reportedly contains rain and low-light splits.","section":"Abstract; §1; §3 (Challenges of Drone Detection)"},{"comment":"The Abstract's 'virtually eliminate motion blur' is a sensor-level statement, but many of the surveyed detectors convert events into fixed-window accumulated frames (Mandula et al. [39], Magrini et al. [35], Zundel et al. [78], etc.). §4 itself notes that frame-based aggregation 'can blur fast motion and miss brief dynamics critical for detecting small, fast drones.' Thus, the systems actually recommended in the survey may not deliver the advertised blur-free operation. Please qualify the claim—native event processing avoids exposure-time blur, but temporal integration into frames reintroduces it—or state which of the reviewed methods preserve true asynchronous processing.","section":"Abstract; §4 (Frame-Based Representations)"}],"minor_comments":[{"comment":"Typo: 'trough' should be 'through' in 'the number of papers trough the years.'","section":"Fig. 1 caption"},{"comment":"The sentence 'Liang et al. Liang et al. [33]' repeats the authors' names. Also, the claim that Magrini et al. [38] is 'the only publicly available benchmark for event-based drone forecasting' is unsupported by a systematic literature check; since [38] is by the same authors, the claim should be softened to 'to our knowledge' or independently verified.","section":"§5 (Drone Forecasting)"},{"comment":"Inconsistent formatting: 'Detr' should be 'DETR'; 'F-UA V-D' and 'Anti-UA V' contain awkward hyphenation/spacing ('UAV' is standard). A style pass would improve readability.","section":"§4 and Table 1"},{"comment":"The header 'RGB / Event' is ambiguous. Adding a short legend (e.g., checkmark means modality is available; otherwise absent) would make the table self-explanatory.","section":"Table 1"},{"comment":"The polarity is defined as pi ∈ {0,1}. Many event-camera papers use {−1,+1}; a brief comment on the convention would avoid confusion.","section":"§3 (formal definitions)"}],"recommendation":"major_revision","confidential_remarks":null},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"You should know two things about this paper up front. First, it's a solid, readable survey of event-based drone detection: it maps representation choices (frames, point clouds, voxels, time-retaining, neuromorphic endpoints), covers detection, tracking, forecasting, and propeller analysis, and includes a handy dataset table. For someone entering the area, it's a good starting point. Second, it has a real internal tension about extreme lighting. The abstract and intro claim that event cameras enable \"consistent detection in extreme lighting\" because of 120 dB dynamic range, but Section 3, in its own list of challenges, says low-light environments and rain/snow \"are challenging for event cameras too\" and cites [9,38]. Low light is an extreme lighting condition. The HDR argument handles high dynamic range within a scene, not low absolute photon flux. The paper never reports stratified results under degraded conditions, so the unqualified claim is not supported by the survey itself. The stress-test note is right about this.\n\nWhat the paper does well: the taxonomies and formal definitions are clean, especially the distinction between drone detection and drone fencing. The treatment of propeller blade analysis is a genuinely useful contribution to the survey—those methods are scattered and the paper ties them together nicely. The dataset table is informative, and the discussion of simulators and synthetic data is fair about domain shift. The inclusion of end-to-end neuromorphic systems, including low-power hardware, gives a practical angle.\n\nThe soft spots, in order of significance. First, the extreme lighting overclaim is the main one; it should be qualified in the abstract and intro, and the Section 3 caveat should be acknowledged there. Second, the claim in Section 5 that Magrini et al. [38] propose \"the only publicly available benchmark for event-based drone forecasting\" is not verified in the paper. It may be true, but for a survey that's a strong statement that needs checking. Third, the authors lean on their own datasets (FRED, NeRDD, Ev-Flying) quite a bit. That's not disqualifying—those datasets are real and published elsewhere—but it does shape the narrative, and the paper doesn't hedge much on that.\n\nWho gets value: newcomers looking for an overview, and researchers choosing sensors or benchmarks. It is a review, not a new result, so don't expect novelty. On the merits, it deserves a serious referee; with minor revision to fix the overclaim and verify the uniqueness claim, it would be a solid reference.\n\nI'd say yes to peer review. The core survey is sound, and the issues are correctable.","headline":"A useful, well-organized survey of event-based drone detection, but its abstract oversells robustness in extreme lighting in a way the paper's own Section 3 contradicts.","tokens_in":803,"tokens_out":977,"would_cite":true,"duration_ms":29791,"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 survey argues that event cameras, by reporting asynchronous brightness changes at microsecond resolution, give counter-UAV systems a sensing foundation frame-based cameras lack.","keywords":["event cameras","neuromorphic vision","drone detection","counter-UAV systems","spiking neural networks","event representations","drone tracking","propeller signature analysis"],"falsifier":"Run a paired field test under the exact failure conditions named in the survey: a small quadrotor flying against a bright sky, in rain, and at decreasing pixel sizes, recorded simultaneously by a frame camera and an event camera. Measure detection rate and latency at each condition. If event-based detection does not beat frame-based detection in the blur or high-contrast regimes, or collapses for very small or distant targets, the survey's central recommendation loses its empirical base.","tokens_in":14856,"feed_emoji":"🛸","tokens_out":7446,"duration_ms":81448,"temperature":0.7,"pith_summary":"The paper is a survey that argues event-based vision is the right sensor foundation for counter-UAV systems because event cameras report per-pixel brightness changes asynchronously at microsecond timescales instead of capturing fixed-rate frames. That single mechanism yields three properties that matter for drone detection: no motion blur for fast targets, a dynamic range above 120 dB that survives backlight and low light, and sparse output that naturally cancels static background. The survey organizes the field by how the event stream is represented—accumulated frames, point clouds, voxel grids, time-retaining frames, or raw spikes for spiking networks—and extends the argument beyond detection to tracking, trajectory forecasting, and identification via propeller blade signatures. A sympathetic reading is that the evidence across datasets and tasks supports the conclusion that event-based vision is a reliable, low-latency, efficient basis for next-generation counter-UAV systems.","feed_headline":"Event cameras can make drone detection blur-free and lighting-proof","feed_subtitle":"Survey shows microsecond sensing and sparse output unlock tracking, forecasting, and propeller-based ID.","key_machinery":"The central object is the event stream $E = \\{e_i = (x_i, y_i, p_i, t_i)\\}$, where each event records pixel coordinates, polarity of the log-brightness change, and a microsecond-resolution timestamp. The argument is carried by the choice of how this asynchronous stream is turned into something a network can consume: accumulated event frames (two polarity channels), 3D voxel grids with time bins, time-retaining frames such as Time Surfaces, event point clouds, or raw spikes for spiking neural networks. Each representation trades retained temporal detail against compatibility with conventional architectures; the survey uses this taxonomy to explain both why early systems work and why end-to-en","core_discovery":"On the paper's own terms, the central claim is that event cameras remove the two failure modes that make frame-based cameras inadequate for drone detection—motion blur from fast angular motion and loss of contrast in extreme lighting—and at the same time provide a built-in attentional mechanism, because static scene elements produce no events. The survey asserts that this combination makes event-based vision a viable foundation for reliable, low-latency counter-UAV systems, and it supports that assertion by tracing the full pipeline: event generation, event representation, detection architectures, benchmarks, and downstream tasks. A distinctive sub-claim is that a drone's propeller, which is","pith_inferences":["Beyond the paper: the same event stream that supports detection could be fused with radar or RF feeds to cover the sensor's known blind spot for stationary drones, since static objects generate no events; a multimodal sensor suite is a natural completion of the counter-UAV picture.","Beyond the paper: the survey's evidence points to a measurable crossover point—at some combination of target angular velocity and apparent size, frame-based detectors degrade while event-based detectors hold; identifying that crossover experimentally would turn the survey's qualitative argument into a design rule.","Beyond the paper: propeller-signature methods reported here imply event cameras could perform long-range, non-cooperative drone identification; the open question is the maximum distance at which propeller events remain resolvable, which would determine whether the method is limited to close-range or scales to surveillance.","Beyond the paper: because the survey reports event-based models outperforming RGB ones on forecast benchmarks, the next test is ablating the event representation itself—frames versus voxel grids versus point clouds versus spiking input—on the same forecasting task to find which encoding carries the advantage."],"forward_implications":["Detection systems can be built around event cameras for scenarios where frame-based cameras saturate or blur, including drones against bright sky, in low light, and in fast close-range maneuvers.","A single event stream can support the full counter-UAV chain—detection, tracking, trajectory forecasting, and identification—so downstream tasks need not require a separate sensor modality.","Propeller signature analysis gives an uncooperative way to identify drones and estimate rotor speed and attitude, enabling applications like virtual fences, autonomous following, and mid-air landing.","Event-plus-RGB multimodal systems can cover the event camera's blind spot for stationary objects while retaining event-based temporal precision.","Low-power neuromorphic processors running spiking networks can move drone detection to the edge with latency and energy budgets that frame-based deep learning cannot match."],"supporting_citations":[{"why":"Supplies the foundational account of event camera principles: asynchronous, microsecond-timestamped brightness-change events and claimed high dynamic range that the entire survey builds on.","marker":"[19]"},{"why":"Provides FRED, the largest multimodal event-RGB drone benchmark with detection, tracking, and forecasting annotations used to compare methods.","marker":"[38]"},{"why":"Introduces NeRDD and a multimodal event-RGB detection pipeline, a central example of frame-based event representations.","marker":"[35]"},{"why":"Provides Ev-UAV, a tiny-object event dataset, and the sparse point-cloud EV-SpSegNet baseline for small or distant drone detection.","marker":"[9]"},{"why":"Introduces EVPropNet, the synthetic-to-real propeller signature detector that establishes propeller analysis as an identification mechanism.","marker":"[52]"},{"why":"Provides the EED event dataset and motion-segmentation approach that first treated drone tracking from event streams.","marker":"[42]"},{"why":"Supplies EventVOT, a high-resolution tracking benchmark including UAVs, and the voxel-based HDETrack baseline.","marker":"[62]"},{"why":"Demonstrates low-power end-to-end neuromorphic drone fencing on an embedded event camera with a spiking neural network, supporting edge-deployment claims.","marker":"[17]"},{"why":"Introduces F-UAV-D, an RGB-event drone dataset with accumulated two-channel event frames, an early frame-based detector example.","marker":"[39]"}],"fun_headline_variants":["Event cameras see drones in blur and darkness","Drone detection that ignores background and blur","Event-based vision spots drones despite motion blur","Sparse events make drone detection fast and robust","Event cameras: no blur, no lighting limits for drones"],"cache_read_input_tokens":2816,"weakest_assumption_plain":"The argument depends on event cameras actually delivering their claimed microsecond temporal resolution, over-120 dB dynamic range, blur-free output, and sparse background suppression in the real conditions where drones must be detected—not just in the controlled or simulated settings cited.","fun_headline_variants_meta":{"raw":{"variants":["Event cameras see drones in blur and darkness","Drone detection that ignores background and blur","Event-based vision spots drones despite motion blur","Sparse events make drone detection fast and robust","Event cameras: no blur, no lighting limits for drones"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000125,"raw_usage":{"total_tokens":911,"prompt_tokens":678,"completion_tokens":233,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":422,"completion_tokens_details":{"reasoning_tokens":163}},"tokens_in":422,"tokens_out":233,"duration_ms":3322,"temperature":1.0,"reasoning_tokens":163,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-05T23:52:13.681274+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run a paired field test under the exact failure conditions named in the survey: a small quadrotor flying against a bright sky, in rain, and at decreasing pixel sizes, recorded simultaneously by a frame camera and an event camera. Measure detection rate and latency at each condition. If event-based detection does not beat frame-based detection in the blur or high-contrast regimes, or collapses for very small or distant targets, the survey's central recommendation loses its empirical base.","supporting_citations":[{"cited_title":"Event-based vision: A survey","cited_arxiv_id":null,"evidence_quote":"Supplies the foundational account of event camera principles: asynchronous, microsecond-timestamped brightness-change events and claimed high dynamic range that the entire survey builds on."},{"cited_title":"Fred: The florence rgb-event drone dataset","cited_arxiv_id":null,"evidence_quote":"Provides FRED, the largest multimodal event-RGB drone benchmark with detection, tracking, and forecasting annotations used to compare methods."},{"cited_title":"Neuromorphic drone detec- tion: an event-rgb multimodal approach","cited_arxiv_id":null,"evidence_quote":"Introduces NeRDD and a multimodal event-RGB detection pipeline, a central example of frame-based event representations."},{"cited_title":"Event-based moving object detection and tracking","cited_arxiv_id":null,"evidence_quote":"Provides the EED event dataset and motion-segmentation approach that first treated drone tracking from event streams."},{"cited_title":"Event stream-based visual object tracking: A high-resolution benchmark dataset and a novel baseline","cited_arxiv_id":null,"evidence_quote":"Supplies EventVOT, a high-resolution tracking benchmark including UAVs, and the voxel-based HDETrack baseline."},{"cited_title":"Drone detection us- ing a low-power neuromorphic virtual tripwire","cited_arxiv_id":null,"evidence_quote":"Demonstrates low-power end-to-end neuromorphic drone fencing on an embedded event camera with a spiking neural network, supporting edge-deployment claims."},{"cited_title":"Towards real-time fast unmanned aerial vehicle de- tection using dynamic vision sensors","cited_arxiv_id":null,"evidence_quote":"Introduces F-UAV-D, an RGB-event drone dataset with accumulated two-channel event frames, an early frame-based detector example."}],"review_version":1}