{"id":"4abaafa4-6b3c-4980-a37d-4f5272008708","arxiv_id":"2506.19416","paper_version":2,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":5,"one_line_summary":"A training-free event-camera pipeline detects MAVs from moving cameras by exploiting the periodic spiral event patterns of fast-rotating propellers, reaching 83.0% precision and 81.5% recall on a new three-drone dataset.","lead":"This paper introduces a training-free method to detect small drones (MAVs) from event cameras by recognizing the swirling patterns their fast-spinning propellers leave in the stream of motion events. It also releases the first event-camera MAV dataset with three drone types and reports higher precision and recall than prior event-based detectors.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The saliency module in Section IV-A assumes positive and negative propeller events overlap within each ΔT/n slice, but n is not reported and the overlap is never validated; if this fails at the tested operating points the pipeline has no input.","rationale":"The reader's weakest_assumption is the same one I find load-bearing, and I agree with it. My reading of rotor physics does not make the overlap impossible: at 10k RPM a blade edge crosses a pixel quickly enough that both polarities can appear within a few milliseconds, especially with multiple blades. But the paper offers no direct measurement or simulation, and the missing parameter n makes the claim untestable from the preprint. The dataset and ablations are genuine contributions, and the reported gains over the baselines are large, so I do not think the paper should be rejected. The correct disposition is the reader's CONDITIONAL, with the condition being an explicit event-level validation of the saliency overlap and a reported n. I set verdict_should_be to UNCHANGED because the reader's verdict already captures this condition; no stronger objection has surfaced after re-reading the method and experiments.","tokens_in":11143,"tokens_out":7624,"duration_ms":93054,"concrete_test":"Use a calibrated event simulator (ESIM or vid2e) to render a two-blade DJI-style propeller at 5k, 10k, and 15k RPM with a measured contrast threshold, over ΔT = 10/15/20/30 ms and n spanning 2-20. For each configuration compute the per-slice overlap ratio |I_p^t ∩ I_n^t| / |I_p^t ∪ I_n^t| in the propeller disk and the maximum accumulated saliency at the propeller center. If no tested (n, ΔT, RPM) combination produces a connected saliency region exceeding τs=50, the core premise of Section IV-A fails; if it only works for one n value, the paper must report n and show the result is stable across it.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim depends on Section IV-A: within each short slice ΔT/n, the binary maps I_p^t and I_n^t of positive and negative events must overlap, giving I_s^t = I_p^t ∩ I_n^t, and this overlap must re-form every rotation cycle so that accumulating over t yields a propeller saliency peak. This is a physical claim about event-camera contrast thresholds, blade geometry, illumination, and slice length, not a mathematical consequence of rotation. The paper never reports n, gives no event-level visualization that separates the two polarities, and provides no simulation or controlled measurement of the overlap fraction. At the cited speeds (5k-15k RPM), one rotation lasts 4-12 ms, so a 10-20 ms period divided into even a few slices can place opposite-polarity transitions in different slices unless blade dwell time and contrast threshold are favorable. If the intersection is empty or too sparse to reach τs=50, the later periodicity and clustering modules have no reliable candidates. The robustness table is consistent with this risk: tiny targets give 39.7% mAP and indoor illumination gives 57.0% mAP, showing the saliency cue is already marginal exactly when event density drops.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes EvDetMAV, a training-free event-camera pipeline for micro aerial vehicle (MAV) detection, built on the observation that rotating propellers produce periodic positive and negative event patterns that are common across different MAV types. The method consists of three modules: a density-aware saliency map that accumulates intersections of positive and negative event maps over slices of the event stream, a spatio-temporal feature extraction stage that scores candidate regions by periodicity, and a clustering-based coarse-to-fine module that outputs bounding boxes. The authors also introduce EventMAV, a new event-based MAV dataset with 25,335 event periods across three MAV types and multiple indoor/outdoor environments, and report state-of-the-art results (83.0% precision, 81.5% recall, F1 82.2%) without any training, outperforming EvPropNet and EV-Tach. The paper claims the first event-based MAV dataset and the first method to handle moving-camera scenarios in this setting.","tokens_in":11273,"tokens_out":4663,"duration_ms":48114,"significance":"If the results hold, the paper makes a useful contribution to event-based MAV detection: it identifies a physically motivated, appearance-free cue (propeller periodicity in the event stream), demonstrates its viability in the previously underexplored moving-camera setting, and provides a new public dataset and codebase. The ablation study (Table VI) and robustness analysis (Table V) are valuable, and the training-free nature is a practical strength, as is the reported real-time operation. The main significance is conditional on the core assumption in Section IV-A being valid and on the method being robust to its hyperparameters; the current manuscript does not yet establish either point with sufficient evidence.","major_comments":[{"comment":"The saliency map is defined by accumulating I_s^t = I_p^t ∩ I_n^t over n slices, so the entire pipeline depends on positive and negative propeller events overlapping in the same pixels within each ΔT/n interval and re-forming every rotation cycle. The paper never reports n, never shows the positive and negative event maps separately, and provides no measurement of the overlap fraction. At the cited 5k–15k RPM, one rotation lasts 4–12 ms, while the dataset periods are 10–30 ms; depending on n, opposite-polarity transitions can easily fall into different slices unless the slice length and contrast thresholds are favorable. Please report n and provide controlled evidence (simulation or event-level statistics) that the intersection is nonempty and dense enough to exceed τ_s=50. This is especially urgent given that the robustness table (Table V) already shows mAP dropping to 39.7% for tiny targets and 57.0% indoors, which is consistent with the saliency cue being marginal exactly when event density is low.","section":"Section IV-A (Fig. 4)"},{"comment":"The periodicity score s_p is described only as counting whether fd, fs, and fp 'have peaks and valleys' after moving-average filtering, with a maximum value of 6, but the peak/valley criterion, the moving-average window, the slice count m, and the scoring formula are not defined. Without these definitions the method is not reproducible, and the threshold τ_p=3 in Algorithm 1 cannot be interpreted or checked. Please specify precisely how peaks and valleys are detected, how the six-point score is accumulated, and what values of m and the filtering window are used.","section":"Section IV-B2 and Algorithm 1"},{"comment":"The method relies on several hand-set thresholds—τ_s=50, τ_p=3, K=4, IoU=0.4—plus n, m, and the peak/valley detection parameters, but no sensitivity analysis, confidence intervals, or error bars are reported. Because these thresholds appear to be chosen on a single self-collected dataset, the generality claim would be much stronger with a sensitivity study showing, for example, that precision/recall are stable over a plausible range of τ_s and τ_p, and that the choice K=4 is not load-bearing. The discussion in Section V-F acknowledges K's role but does not quantify the effect of varying it.","section":"Section V-A and V-D"},{"comment":"The evaluation is conducted entirely on the authors' own dataset, with manually annotated bounding boxes and IoU=0.4, and no error bars or per-sequence variability are reported. While this is understandable given the absence of other event-based MAV datasets, the headline numbers should be accompanied by a statement of how the thresholds were selected (training set, validation set, or test set) and by some measure of variance, so that the reader can judge whether the +30.3%/+36.4% margins over baselines are meaningful rather than the result of parameter tuning on the test distribution.","section":"Section V-B (Table III)"}],"minor_comments":[{"comment":"The abstract reports a recall improvement of +36.4% over EvPropNet, but Section V-B says +34.6%; the arithmetic 81.5−45.1 = 36.4 supports the abstract, so the number in Section V-B should be corrected.","section":"Abstract and Section V-B"},{"comment":"The sentence 'The largest eigenvalue λ_max and its corresponding eigenvector ξ has the relation λ_max C = ξ C' appears to be a typo; it should state C ξ = λ_max ξ. Please clarify.","section":"Section IV-B1, text near Eq. (1)"},{"comment":"The phrase 'Calculate the Gaussian shape' is never defined in the text; the fine stage needs a precise description of what Gaussian fitting is performed and how 'consistent with P_i' is checked.","section":"Algorithm 1, fine stage"},{"comment":"The abbreviation 'SSN' is used where 'SNN' (spiking neural network) is meant; this should be fixed throughout the paragraph.","section":"Section II-C"},{"comment":"The caption states that the event images of EvPropNet contain a shorter period than the other two methods; this difference should be stated in the main text and the exact periods used for each method should be reported so that the qualitative comparison is fair and reproducible.","section":"Figure 6 caption"}],"recommendation":"major_revision","confidential_remarks":"The paper addresses a timely problem and the dataset is a potentially valuable community resource. The main risk is that the core saliency assumption in Section IV-A is not independently validated and the method's many thresholds are not analyzed for sensitivity. I would ask the authors to add the missing validation and sensitivity experiments before acceptance; the central idea is defensible and the revision is feasible within the manuscript's scope."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"EvDetMAV is a solid engineering paper with a genuinely new idea and a useful new dataset. The core trick—slicing event data, intersecting positive and negative event images per slice, and accumulating those intersections to highlight spinning propellers—is not something I've seen in the event-camera detection literature. It's simple, physically motivated, and apparently effective: 83.0% precision and 81.5% recall on their test set, far ahead of EvPropNet and EV-Tach, and the ablations show each module contributes. The EventMAV dataset (25k annotated periods, three DJI platforms, moving camera, varied environments) is a resource the community will likely use.\n\nThe soft spots are real but not fatal. The method has several hand-set thresholds (τ_s=50, τ_p=3, K=4) with no sensitivity analysis or error bars. The periodicity score relies on a peak/valley detector described only loosely. The slice count n in the saliency module isn't reported, and the whole method depends on positive and negative events overlapping within each slice. The method's success on the dataset suggests that assumption often holds, but the authors don't validate it directly. The robustness table shows performance collapsing for tiny targets (39.7% mAP) and indoor scenes (57.0% mAP), consistent with the saliency cue being marginal when event density drops.\n\nAlso, 'generalized MAV detection' is tested on three DJI quadrotors with exposed propellers, one target per period. That's a reasonable start, but it doesn't cover enclosed propellers (the paper does note that limitation) or multi-target scenes. There's a minor discrepancy in the reported recall improvement—+36.4% in the abstract, +34.6% in the body. Not a big deal, but should be fixed.\n\nOverall, this is a contribution worth engaging with. The dataset alone justifies the paper, and the method is a creative addition. Before I'd fully trust the quantitative claims, the authors should provide threshold sweeps, error bars, the slice count n, and a more detailed description of the periodicity features. It deserves a serious referee, and I'd bring it to the reading group.\n\nRecommendation: send to peer review, with requests for revision.","headline":"Clever, training-free drone detection from propeller event patterns, backed by a useful new dataset—but the generalization claim outruns the evidence.","tokens_in":11946,"tokens_out":3476,"would_cite":true,"duration_ms":35214,"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":"Different drone types share a propeller event signature that lets a training-free pipeline detect them from moving event cameras.","keywords":["event camera","MAV detection","propeller detection","saliency map","periodicity","training-free detection","event-based vision","micro aerial vehicle"],"falsifier":"Record the event stream of a single spinning propeller whose blades trigger only one contrast polarity (for example, a blade with one reflective surface under steady illumination) and check whether the positive and negative event regions overlap within any 1/n slice of a rotation; if they never overlap, the density-aware saliency map cannot highlight the propeller and detection fails.","tokens_in":10816,"feed_emoji":"🚁","tokens_out":6413,"duration_ms":63786,"temperature":0.7,"pith_summary":"Micro aerial vehicles of different makes and models share one feature that ordinary RGB images hide: fast-spinning propellers produce dense, periodic streams of positive and negative events in an event camera. This paper argues that this shared physical signature is enough for generalized MAV detection, and builds EvDetMAV, a training-free pipeline that detects propellers directly from raw event streams even when the camera itself is moving. The pipeline generates a density-aware saliency map by intersecting positive and negative event regions within short rotation slices, then scores candidate regions by periodicity and refines them with clustering. On the new EventMAV dataset of 25,335 event periods from three MAV types, the method reports 83.0% precision and 81.5% recall without any training, a gain of roughly 30 points over prior event-based detectors. The paper also releases the first moving-camera, multi-type event-based MAV dataset for the community.","feed_headline":"Spinning propellers make drones visible to moving event cameras","feed_subtitle":"A training-free pipeline hits 83% precision and 81.5% recall on three drone types in moving-camera scenes.","key_machinery":"The carrying mechanism is a polarity-intersection saliency map. For a propeller spinning at thousands of RPM, each point on the blade sweeps through the same image location twice per revolution with opposite contrast, so within a short sub-interval $\\frac{1}{n}\\Delta T$ the positive-event image $I_p^t$ and negative-event image $I_n^t$ overlap in the propeller region; their intersection $I_s^t = I_p^t \\cap I_n^t$ is nonempty and re-accumulates each rotation cycle. Summing $I_s^t$ over $n$ slices builds a saliency map in which the periodic propeller signature is enhanced while events from background objects and camera motion, which do not revisit the same pixels with alternating polarity, average out. The same periodicity is then quantified per candidate region through three descriptors, density fluctuation, structural similarity of consecutive slice images, and principal-direction angle change, and used with a coarse-to-fine clustering step to produce the final bounding box.","core_discovery":"The paper's central claim is that fast-rotating MAV propellers leave a distinctive, appearance-independent trace in the raw event stream: a periodic alternation of positive and negative events that survives camera motion and background clutter, and that this trace can be exploited without learned features. In the proposed pipeline, each short period $\\Delta T$ is split into $n$ slices; in each slice the binary positive image $I_p^t$ and negative image $I_n^t$ are intersected to form $I_s^t$, based on the observation that a rotating blade passes the same pixels with opposite contrast within a rotation. Accumulating these intersections over the period suppresses background noise produced by camera motion and yields a saliency map. Candidate areas are then scored by saliency and by periodicity descriptors (event density, structural similarity between consecutive slices, and principal-direction similarity of the point cloud), and a coarse-to-fine clustering stage outputs the bounding box. Without any training, the method reaches 83.0% precision and 81.5% recall on the proposed EventMAV testing set, outperforming EvPropNet by +30.3 precision and +36.4 recall and EV-Tach by similar margins, and also exceeding CNN-based baselines trained on other MAV subsets.","pith_inferences":["Beyond the paper's claims, the underlying signature is purely physical, so the same training-free pipeline should transfer to any rotorcraft with exposed rotating blades, such as helicopters or fixed-wing propellers, provided the rotation rate is high enough to produce alternating-polarity events within one slice.","Beyond the paper's claims, the method's failure modes suggest a natural fusion: a learned RGB or IR detector could supply appearance context at long range and close to the ground, while the propeller-periodicity detector covers unseen platforms and fast camera motion.","Beyond the paper's claims, the periodicity scores could be repurposed as a weak rotation-rate estimator, since the peak spacing in density and structural-similarity signals is tied to blade-passing frequency; this would make the saliency map a byproduct of tachometry rather than a separate stage."],"forward_implications":["An observer using a moving event camera can detect drones of types it has never seen, as long as their propellers are exposed and spinning.","The EventMAV dataset provides the first benchmark for this task, with 25,335 annotated event periods covering three MAV types, 15 environments, and multiple viewing angles and scales.","The full pipeline runs in about 20 ms per period on a CPU, so real-time detection on resource-limited platforms is plausible.","Performance degrades gracefully under hard conditions: it drops to 39.7% mAP for tiny targets and is worse indoors than outdoors, while large, close, vertical views achieve over 80% mAP.","Combining the propeller features with learned networks is an explicit direction for further improvement suggested by the paper."],"supporting_citations":[{"why":"It supplies the physical premise that MAV propellers spin at 5k to 15k RPM.","marker":"[7]"},{"why":"It is the prior event-propeller detection method whose reported performance the proposed method surpasses.","marker":"[9]"},{"why":"It is the prior event-based rotational-speed method used as a comparison baseline.","marker":"[10]"},{"why":"It provides the prior event-camera MAV detection setup limited to stationary cameras and the source of the YOLO comparison configuration.","marker":"[11]"},{"why":"It supplies the CNN detector used for the generalization comparison in the experiments.","marker":"[17]"},{"why":"It establishes the domain-gap problem in RGB MAV detection that motivates generalized event-based detection.","marker":"[6]"}],"fun_headline_variants":["Drone propellers leave a unique event-camera signature","No training needed: event-based drone detection from propellers","Event cameras spot drones by their spinning propellers","Moving event cameras detect MAVs via propeller flicker","Propeller trails power training-free drone detection in events"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The saliency map assumes that a spinning propeller repeatedly triggers both brightening and darkening events at the same set of pixels within each short slice of time.","fun_headline_variants_meta":{"raw":{"variants":["Drone propellers leave a unique event-camera signature","No training needed: event-based drone detection from propellers","Event cameras spot drones by their spinning propellers","Moving event cameras detect MAVs via propeller flicker","Propeller trails power training-free drone detection in events"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000506,"raw_usage":{"total_tokens":2508,"prompt_tokens":1025,"completion_tokens":1483,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":641,"completion_tokens_details":{"reasoning_tokens":1406}},"tokens_in":641,"tokens_out":1483,"duration_ms":10117,"temperature":1.0,"reasoning_tokens":1406,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-06T23:07:05.421716+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Record the event stream of a single spinning propeller whose blades trigger only one contrast polarity (for example, a blade with one reflective surface under steady illumination) and check whether the positive and negative event regions overlap within any 1/n slice of a rotation; if they never overlap, the density-aware saliency map cannot highlight the propeller and detection fails.","supporting_citations":[{"cited_title":"Quan, Introduction to Multicopter Design and Control","cited_arxiv_id":null,"evidence_quote":"It supplies the physical premise that MAV propellers spin at 5k to 15k RPM."},{"cited_title":"EVPropNet: Detecting drones by finding propellers for mid-air landing and following,","cited_arxiv_id":null,"evidence_quote":"It is the prior event-propeller detection method whose reported performance the proposed method surpasses."},{"cited_title":"EV-Tach: A handheld rotational speed estimation system with event camera,","cited_arxiv_id":null,"evidence_quote":"It is the prior event-based rotational-speed method used as a comparison baseline."},{"cited_title":"Event-based mo- tion capture system for online multi-quadrotor localization and tracking,","cited_arxiv_id":null,"evidence_quote":"It provides the prior event-camera MAV detection setup limited to stationary cameras and the source of the YOLO comparison configuration."},{"cited_title":"YOLOv5 by Ultralytics,","cited_arxiv_id":null,"evidence_quote":"It supplies the CNN detector used for the generalization comparison in the experiments."},{"cited_title":"Domain adaptive detection of MA Vs: A benchmark and noise suppression network,","cited_arxiv_id":null,"evidence_quote":"It establishes the domain-gap problem in RGB MAV detection that motivates generalized event-based detection."}],"review_version":1}