{"id":"1946e84d-c083-454a-91d0-0c667d52b99d","arxiv_id":"2501.10407","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":1,"one_line_summary":"RadDet is a new open-source dataset for wideband radar spectrum detection with 40,000 annotated frames, 11 radar classes, 6 SNR settings, 2 density environments, and 3 time-frequency resolutions.","lead":"This paper introduces RadDet, a synthetic dataset of 40,000 radar spectrum frames with labels that mark where and when radar signals appear across a 500 MHz band. It also benchmarks four real-time object detection models on RadDet and on an existing NIST radar dataset.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Equation (1) defines temporal bbox width as N/fprf, yielding 20–100 s boxes in a 2 ms frame; the annotation formula as written is dimensionally wrong and unverifiable.","rationale":"I read the paper as proposing a synthetic wideband radar detection dataset with precise time-frequency annotations and a benchmark. The load-bearing condition for the central claim is that the annotations are correct. The reader's weakest assumption focused on signal realism; while that is a legitimate transfer concern, a more immediate and concrete problem is the formal definition of the temporal bounding box. Eq. (1) as written cannot produce valid boxes: for RadDet's N=1e6 samples and fprf between 10 and 50 kHz, the width N/fprf is 20–100 s, while the frame duration is N/fs = 2 ms. This is not a subtle regime issue; it is a dimensional error in the published specification. If the released dataset was generated from a corrected version of the formula (e.g., using Npulse instead of N), then the paper merely needs a correction. If it was generated as written, the annotations are nonsense and the benchmark is invalid. The concrete test described above distinguishes these cases. I therefore support the reader's CONDITIONAL verdict, but with the condition elevated from 'clarify the variable' to 'verify and correct Eq. (1), or demonstrate that the released annotations match the intended pulse-train extent.' My read does not move the verdict because the reader already required clarification; it sharpens what is required.","tokens_in":8598,"tokens_out":6282,"duration_ms":57449,"concrete_test":"Download a RadDet annotation file for a frame containing a single pulsed radar (e.g., class Rect or P0N) with known tpw, fprf, and Npulse. Compute the expected pulse-train temporal extent (Npulse-1)/fprf + tpw and compare it to the annotation's t_bbox width. Also check whether any t_bbox width exceeds the 2 ms frame duration (N/fs = 1e6/500e6). Independently, re-implement Eq. (1) for NIST-CBRS with N=800,000 and fprf=10 kHz; if the computed width (80 s) greatly exceeds the signal duration (0.08 s), the formula is not what was implemented and the paper needs a correction.","verdict_should_be":"UNCHANGED","load_bearing_attack":"Section III-A defines the temporal bounding-box width as t_bbox in [ts, ts + N/fprf] (Eq. 1), with N stated as 'the sample length.' For RadDet, N=1,000,000 and fprf ranges over 10–50 kHz (Table I), producing widths of 20–100 seconds, orders of magnitude larger than the 2 ms frame duration (N/fs = 2 ms). Even for NIST-CBRS (N=800,000, fprf=10–50 kHz), Eq. (1) gives 16–80 s boxes for 0.08 s signals. The formula is dimensionally inconsistent (a count divided by a frequency yields seconds times samples) and cannot describe a pulse train's temporal extent, which should be roughly (Npulse-1)/fprf + tpw. Section III-B states annotations are computed per (1) and (2), so either the released dataset uses a different rule or the paper's specification is wrong. Because the central claim is a dataset with correct time-frequency annotations, this internal inconsistency is more load-bearing than the realism assumption: even a perfectly realistic signal generator cannot repair ground-truth boxes whose widths are specified to be roughly 1000 times the frame length. Without code or a corrected formula, Table III benchmark numbers cannot be trusted.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"This paper introduces RadDet, a synthetic open-source wideband radar spectrum detection dataset with 40,000 frames, each derived from 1 million I/Q samples across a 500 MHz band, covering 11 radar classes, 6 SNR settings, sparse and dense environments, and 3 time-frequency resolutions. The authors also adapt the NIST-CBRS dataset into a detection benchmark and evaluate YOLOv3, YOLOv6, YOLOv9, and RT-DETR with mAP and FPS metrics. The central claim is that RadDet provides a large-scale, publicly available benchmark with time-frequency and class annotations for real-time radar spectrum detection, and that real-time detectors can operate on it at hundreds of frames per second.","tokens_in":8860,"tokens_out":6561,"duration_ms":64694,"significance":"If the dataset and annotations are correct and the code and data are actually released, RadDet fills a genuine gap: public wideband radar detection datasets with dense multi-emitter frames, localisation annotations, and controlled SNR and density variations are scarce. The paper's strengths include the large scale, the systematic variation of SNR, density, and time-frequency resolution, the use of standard detection architectures, and the detailed training protocol. However, the annotation formula in Eq. (1) is dimensionally inconsistent, making the correctness of the ground-truth boxes unverifiable; the repository link is a placeholder; and the synthetic-to-real transferability is asserted rather than validated. These issues are load-bearing for the dataset's central contribution and must be resolved before the benchmark can be used as a reference.","major_comments":[{"comment":"Equation (1) specifies the temporal bounding box as [ts, ts + N/fprf], with N stated to be the sample length. For RadDet, N=10^6 and fprf ranges over 10-50 kHz (Table I), so the formula gives widths of 20-100 seconds, whereas each frame is only N/fs = 2 ms long. This is dimensionally inconsistent: a number of samples divided by a frequency does not yield a time interval, and it cannot describe the temporal extent of a pulse train, which should be roughly (Npulse-1)/fprf + tpw. Since Section III-B states that all annotations are computed per Eqs. (1)-(2), either the released dataset was annotated with a different rule or the paper mis-specifies the rule. I request a corrected formula (e.g., using Npulse or an explicit pulse-train duration), release of the annotation-generation code, and confirmation that the Table III results were produced with the corrected annotations.","section":"Section III-A, Eq. (1)"},{"comment":"The paper claims the dataset is open-source, but the URL https://github.com/abcxyzi/RadDet is a placeholder. The central contribution of the paper is the dataset itself, so a working repository containing the full dataset, generation scripts, and annotation code is necessary for the claims to be verifiable. Please provide the actual URL and ensure it is live at the time of publication.","section":"Section I, footnote 1"},{"comment":"The paper asserts that the emitter parameter library from [32] bounds the sampling range and that AWGN models the noise environment, but it provides no comparison with measured radar signals, no channel effects beyond AWGN, and no independent sanity check of the generated signals. Because the utility of the benchmark depends on transfer to real non-cooperative radar environments, the authors should at least state this limitation explicitly and, if possible, include a small validation set of real or independently generated signals to support the realism claim.","section":"Section III-B, emitter parameter realism"}],"minor_comments":[{"comment":"The word 'genreated' should be 'generated'.","section":"Section III-B"},{"comment":"The parameter name 'tprf (kHz)' is inconsistent with Eq. (1)'s 'fprf'; rename it to fprf or PRF for consistency.","section":"Table I"},{"comment":"Equation (1) calls tbbox the bbox width but writes it as an interval; clarify that the interval denotes the time range and that the width is the length of that interval.","section":"Eq. (1)"},{"comment":"The dimensions '1282', '2562', and '5122' appear to be missing superscripts (128^2, 256^2, 512^2); fix the formatting.","section":"Table II"}],"recommendation":"major_revision","confidential_remarks":"The manuscript is a straightforward dataset-and-benchmark paper, and the self-citation to the authors' RadSeg work [20] is reasonable given the direct lineage. My main editorial concern is the placeholder repository and the Eq. (1) error; I would be willing to accept after the authors supply a corrected formula, a working data/code release, and a clear statement of the synthetic-to-real limitations."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Colleague, here's my read on RadDet. The dataset itself is the real contribution: 40k frames, 500 MHz, 11 classes including six polyphase LPI waveforms, dense and sparse scenes, three spectrogram resolutions, YOLO-format boxes. That is a genuine gap. NIST-CBRS is single-instance and narrowband; RadSeg has temporal-only labels. So the resource alone justifies attention. The benchmark is plain-vanilla but fine: four real-time detectors, three input sizes, sensible metrics. The observation that augmentation helps and RT-DETR struggles on small LPI objects is credible.\n\nThe soft spot is the annotation formula. Eq (1) says t_bbox = ts + N/fprf, where N is sample length. For RadDet that is 1e6 / 10-50 kHz = 20-100 s, in a 2 ms frame. That cannot be right. The correct temporal span of a pulse train should involve pulse count and period, not sample count over PRF. The paper then says all annotations are computed from (1)-(2), so either the published formula is a typo and the actual generator uses something else, or the released boxes are nonsense. The reader was right to flag this as load-bearing. For a dataset paper, the annotation spec is the product. You cannot review the benchmark meaningfully until this is resolved and the code is released. The GitHub link is given, but I couldn't verify what is actually there; if the code and annotations are public, the authors can settle this immediately.\n\nOther issues are minor by comparison: no error bars across runs, synthetic data realism rests on sampling ranges from a textbook library, and the max-hold spectrogram preprocessing is a choice that needs justification but is not fatal. The writing is clear and the empirical claims are modest; I saw no self-citation inflation beyond citing their own RadSeg as the natural predecessor.\n\nWho is this for? Anyone working on radar spectrum sensing, cognitive EW, or time-frequency object detection. It deserves a serious referee, but only after the formula is corrected and the artifact is verified. I would send it back for major revision on those points.","headline":"A genuinely useful wideband radar dataset, but the printed annotation formula is dimensionally wrong and that must be fixed before the dataset can be trusted.","tokens_in":9407,"tokens_out":2692,"would_cite":true,"duration_ms":25166,"reading_group":"yes","serious_thinker":"unclear","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"This paper introduces RadDet, a public dataset of 40,000 synthetic wideband radar frames spanning 500 MHz with 11 radar classes, and demonstrates real-time YOLO-family detectors localising radar emissions at hundreds of frames per second.","keywords":["radar spectrum detection","wideband dataset","real-time detection","time-frequency annotations","YOLO","spectrogram","low probability of intercept","electronic warfare"],"falsifier":"Take the trained RadDet detectors and run them on a real recorded wideband radar capture, digitised I/Q from a maritime or airborne setting with operator-confirmed radar annotations, and compare per-class mAP50 to the synthetic test set. If real-world scores fall far outside the 20-60% mAP50 range reported here, the benchmark's realism assumption is falsified.","tokens_in":8406,"feed_emoji":"📡","tokens_out":6027,"duration_ms":53983,"temperature":0.7,"pith_summary":"The paper tries to remove a data bottleneck: radar spectrum detection research lacks public wideband datasets with time-frequency annotations, so models cannot be compared or trained in realistic dense environments. RadDet provides 40,000 frames, each generated from one million I/Q samples across a 500 MHz band, covering 11 radar classes, six SNR settings, two radar-density environments, and three time-frequency resolutions. The authors train standard real-time object detectors on the spectrogram frames and report detection accuracy and speed, showing that the task is feasible at hundreds of frames per second. If the synthetic signals faithfully represent non-cooperative radar environments, RadDet becomes a repeatable benchmark for a capability that previously had no common testbed.","feed_headline":"Wideband radar detection gets a 40,000-frame public benchmark","feed_subtitle":"Synthetic 500 MHz spectrum with 11 radar classes lets YOLO-style detectors run at hundreds of frames per second.","key_machinery":"The engine is the max-hold spectrogram pipeline: each frame's one million complex I/Q samples are transformed with a short-time Fourier transform, and a max-hold pooling step compresses the time axis while preserving bright radar returns. Bounding boxes are computed from the known radar parameters through the paper's equations (1)-(3), so each radar instance gets a time-frequency box in YOLO format. The dataset is offered in sparse (at most one emitter per frame) and dense (up to nine emitters per frame) variants, and the benchmark models are standard real-time detectors (YOLOv3, YOLOv6, YOLOv9, RT-DETR) trained on these spectrogram images.","core_discovery":"The central claim is that RadDet is a wideband radar detection benchmark with bounding-box-level annotations for radar dwells, going beyond narrowband or single-instance datasets. On the dense version, at the finest resolution, the strongest YOLO-family model reaches about 60% mAP50, while the sparse version is much harder, with mAP50 near 20-30%; inference speeds range from roughly 200 to over 1000 frames per second depending on resolution and model. The paper also converts the 3.5 GHz CBRS radar dataset into a detection dataset with computed time-frequency boxes, serving as a baseline. The discovery is that real-time object detection recipes transfer to spectrogram images well enough to localise multiple overlapping radar signals, with augmentation and resolution as the main controls on the accuracy-speed trade-off.","pith_inferences":["My inference: the dataset's realism is the unvalidated link; an obvious next experiment is to record a small set of real over-the-air LPI radar emissions and measure how much the detector's mAP drops on those recordings.","My inference: because frames are synthetic and parameterised, the same generation pipeline could produce unlimited variants, so a follow-up could expand SNR ranges, add interference, or add new classes without recollecting data.","My inference: the bounding-box convention treats each radar dwell as one object, but pulse-level segmentation would be a natural harder task that the same dataset could support if masks were generated.","My inference: the max-hold pooling factor controls the accuracy-speed trade-off; making that factor adaptive per frame or per frequency band might recover accuracy in dense scenes without sacrificing real-time speed."],"forward_implications":["Labs working on radar detection can now compare models on identical frames instead of private datasets.","Deployable wideband monitoring is plausible: at the smallest input resolution, detectors run at over 1000 frames per second while still detecting signals.","Time-frequency resolution is a tunable operating point: moving from 128x128 to 512x512 inputs roughly doubles dense-scene mAP50 but cuts frame rate by about a factor of four.","Sparse radar environments expose a concrete weakness in current detectors: small LPI waveforms and rare pulses keep mAP50 below 35%."],"supporting_citations":[{"why":"Supplies the long-sequence radar construction that RadDet extends, including the notion of radar pulse activity as a detection target.","marker":"[20]"},{"why":"Provides the max-hold spectrogram generation and detection approach adapted for RadDet, including the basis for the bounding-box formulas.","marker":"[29]"},{"why":"Provides the original CBRS radar dataset and class distribution that RadDet mirrors and whose time-frequency boxes are computed.","marker":"[28]"},{"why":"Supplies the emitter parameter library that bounds RadDet's synthetic radar sampling ranges.","marker":"[32]"},{"why":"Defines the YOLO bounding-box annotation format used for all datasets.","marker":"[33]"},{"why":"YOLOv3 is the strongest baseline detector on RadDet and defines the real-time detection paradigm.","marker":"[34]"},{"why":"RT-DETR is the transformer baseline whose poor small-object performance supports the paper's model comparison.","marker":"[37]"},{"why":"RadioML is a representative classification dataset RadDet contrasts against, showing the lack of localisation annotations in prior work.","marker":"[7]"}],"fun_headline_variants":["RadDet: 40K wideband radar frames for real-time detection","Wideband radar detection benchmark: 40K frames, 11 classes","Real-time radar spectrum detection gets a 40,000-frame dataset","YOLO on radar: 40K-frame dataset for wideband detection","New radar dataset enables real-time wideband spectrum detection"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The whole benchmark stands on the premise that the synthetic radar signals, with parameters drawn from an emitter library and noise added as AWGN, are representative enough of real non-cooperative radar environments for measured performance to transfer.","fun_headline_variants_meta":{"raw":{"variants":["RadDet: 40K wideband radar frames for real-time detection","Wideband radar detection benchmark: 40K frames, 11 classes","Real-time radar spectrum detection gets a 40,000-frame dataset","YOLO on radar: 40K-frame dataset for wideband detection","New radar dataset enables real-time wideband spectrum detection"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000312,"raw_usage":{"total_tokens":1759,"prompt_tokens":914,"completion_tokens":845,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":530,"completion_tokens_details":{"reasoning_tokens":752}},"tokens_in":530,"tokens_out":845,"duration_ms":8604,"temperature":1.0,"reasoning_tokens":752,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-10T22:03:50.303157+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Take the trained RadDet detectors and run them on a real recorded wideband radar capture, digitised I/Q from a maritime or airborne setting with operator-confirmed radar annotations, and compare per-class mAP50 to the synthetic test set. If real-world scores fall far outside the 20-60% mAP50 range reported here, the benchmark's realism assumption is falsified.","supporting_citations":[{"cited_title":"Multi-Stage Learning for Radar Pulse Activity Segmentation,","cited_arxiv_id":null,"evidence_quote":"Supplies the long-sequence radar construction that RadDet extends, including the notion of radar pulse activity as a detection target."},{"cited_title":"Deep Learning for Radar Signal Detection in the 3.5 GHz CBRS Band,","cited_arxiv_id":null,"evidence_quote":"Provides the max-hold spectrogram generation and detection approach adapted for RadDet, including the basis for the bounding-box formulas."},{"cited_title":"RF Dataset of Incumbent Radar Signals in the 3.5 GHz CBRS Band,","cited_arxiv_id":null,"evidence_quote":"Provides the original CBRS radar dataset and class distribution that RadDet mirrors and whose time-frequency boxes are computed."},{"cited_title":"Lynch Jr., An Introduction to RF Stealth","cited_arxiv_id":null,"evidence_quote":"Supplies the emitter parameter library that bounds RadDet's synthetic radar sampling ranges."},{"cited_title":"You Only Look Once: Unified, Real-Time Object Detection,","cited_arxiv_id":null,"evidence_quote":"Defines the YOLO bounding-box annotation format used for all datasets."},{"cited_title":"DETRs Beat YOLOs on Real- Time Object Detection,","cited_arxiv_id":null,"evidence_quote":"RT-DETR is the transformer baseline whose poor small-object performance supports the paper's model comparison."},{"cited_title":"Over-the- Air Deep Learning Based Radio Signal Classification,","cited_arxiv_id":null,"evidence_quote":"RadioML is a representative classification dataset RadDet contrasts against, showing the lack of localisation annotations in prior work."}],"review_version":1}