{"id":"9ee41ba0-c23c-486f-950a-edffb3a01e3b","arxiv_id":"2508.03331","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":0,"one_line_summary":"LRDDv2 is a new public dataset with 39,516 annotated drone images, range labels on over 8,000 images, and benchmarks showing better YOLOv8 detection than training on Drone-vs-Bird alone.","lead":"This paper releases LRDDv2, a dataset of 39,516 annotated images of distant drones, with range information for over 8,000 of them. It is built for researchers working on detecting small drones at long range, a key safety need as drones become common.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Range labels for over 8,000 images lack any described measurement method or validation; without a public error analysis, the dataset's headline feature and any range-estimation benchmark are unsupported.","rationale":"The reader identified the accuracy and provenance of the range information as the weakest assumption, and this is also the most load-bearing concern in my reading. The abstract and conclusion explicitly market the range labels as the dataset's standout feature; the rest of the contribution is a straightforward, though useful, collection of annotated drone images. If the range labels are unreliable, the primary scientific novelty collapses, and the benchmark results, though interesting, do not rescue the paper because they measure detection accuracy, not range-estimation accuracy. I therefore agree with the reader's conditional verdict: the paper should be accepted only if the authors provide a clear description of the range measurement procedure and a quantitative validation of label accuracy. I considered whether the benchmark results without statistical significance are equally load-bearing, but those results are supporting evidence for dataset utility, not the central claim; the central claim is the dataset itself. No adjustment to the reader's verdict is needed, only the requested conditions should be attached to acceptance.","tokens_in":4250,"tokens_out":2873,"duration_ms":35383,"concrete_test":"Select a random subset of 200 images from the over-8,000 range-labeled images. For each image, independently obtain the true distance using a synchronized, high-accuracy method such as RTK GPS logs from the drone or a laser rangefinder at capture time, and compare these measured distances with the dataset's labels. Report the mean absolute error, median relative error, and Pearson correlation between labels and ground truth. If the median relative error exceeds 10% or the correlation falls below 0.95, the range-information claim is not supported and the conditional acceptance should be revisited.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The paper's central novelty is the inclusion of range information for over 8,000 images, stated in the abstract and reiterated in Section 8 as the 'standout feature.' However, the visible manuscript contains no description of how these ranges were measured (e.g., GPS/RTK, laser rangefinder, structure-from-motion, or estimation from known drone size and focal length), no statement of which camera or capture geometry was used, and no validation of label accuracy. If the range labels are inaccurate, noisy, or systematically biased, models trained on them cannot learn a reliable image-to-range mapping, and the claimed enabling capability for range estimation is compromised. This is load-bearing because the dataset's distinct value over existing drone detection datasets rests on that range information; without it, LRDDv2 becomes primarily another small-object detection set. The absence of validation also weakens the paper's downstream claims about detection probability versus bounding-box area, since bounding-box area is treated as a proxy for range. The reader's conditional verdict correctly targets this gap, but the concern is not merely missing documentation: it is an unverified scientific claim about label quality that must be tested and reported.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper introduces LRDDv2, a second release of the Long-Range Drone Detection dataset, comprising 39,516 annotated images with a stated emphasis on long-range small-object detection and with target range information claimed for over 8,000 images. The authors report YOLOv8 detection experiments showing that models trained on LRDDv2 outperform models trained solely on the Drone-vs-Bird dataset when evaluated on the external Detfly and UAV-Detect datasets, and that combining both training sets yields the best results. The paper also analyzes detection probability as a function of bounding-box area and proposes future work on range estimation enabled by the range labels.","tokens_in":4449,"tokens_out":1801,"duration_ms":24889,"significance":"If the range information is accurate and the dataset is released as described, LRDDv2 would be a useful contribution: existing drone-detection datasets rarely include range labels, and the empirical comparison on external evaluation sets (Detfly, UAV-Detect) gives concrete evidence about the value of diverse long-range training data. The dataset size alone is substantial for this niche. However, the central novelty—the range labels—is currently an unsupported claim: the manuscript documents neither how the ranges were measured nor how their accuracy was checked. Because the paper's distinguishing value over prior drone datasets rests on these labels, the contribution cannot be fully assessed from the preprint. The benchmark results are presented without training details or variability measures, which limits reproducibility but does not invalidate the qualitative trend.","major_comments":[{"comment":"The paper's headline feature—explicit target range information for over 8,000 images—is not accompanied by any description of how those ranges were measured. The manuscript never states whether distances came from GPS/RTK logs, a laser rangefinder, known drone size projected through the camera model, or another method, nor does it report the camera(s) and capture geometry used. Without this methodology, the range labels are unverifiable, and the claimed capability of training range-estimation algorithms is not supported.","section":"Abstract and Section 8"},{"comment":"No validation or error analysis of the range labels is provided. There is no statement about the accuracy tolerance, no comparison against a ground-truth measurement, and no per-image confidence or flags for uncertain ranges. Because the range label is the primary novelty, the authors should report a quantitative validation (e.g., mean absolute error against independent measurements, or inter-annotator consistency) and state the distribution of ranges across the 8,000+ images.","section":"Section 8 (dataset description)"},{"comment":"The analysis of detection probability versus bounding-box area treats bounding-box area as a direct proxy for range ('As the distance increases (represented by smaller bounding boxes)'), yet the paper never calibrates bounding-box area to actual distance using the range labels that it claims to provide. Given that range information exists for over 8,000 images, the authors should either plot detection probability against the measured range or explicitly justify why bounding-box area is an adequate substitute, including its variability at fixed range due to differing drone sizes and camera focal lengths.","section":"Section 6"},{"comment":"The reported benchmark lacks essential experimental details needed to assess or reproduce the comparison: the number of training epochs, hyperparameters, input resolution, data split between training and validation, and whether the evaluation sets were used only for testing are not specified. Additionally, only single-point mAP values are given without variance across runs. Since the paper's main empirical claim is that LRDDv2 training improves detection on Detfly and UAV-Detect, these details and some measure of run-to-run variability are needed.","section":"Table 2 and Section 5 (benchmark)"}],"minor_comments":[{"comment":"There are numerous typos and formatting issues, including 'UA Vs' with an inconsistent space, 'Drone vs Birds' inconsistent capitalization, and a stray sentence fragment '0.32. The improvements represent...' before Table 2. These should be corrected.","section":"Throughout"},{"comment":"The paper should include a summary table of dataset statistics—number of images per category (drone vs. no-drone, if applicable), number of bounding boxes, image resolution distribution, distribution of bounding-box sizes, and environmental conditions—so readers can evaluate the dataset's composition without downloading it.","section":"Section 1 / Dataset description"},{"comment":"The conclusion states 'over 39,516' images, while the abstract and introduction state '39,516' exactly; the discrepancy should be resolved.","section":"Section 8"},{"comment":"Reference [14] is the authors' own LRDDv1 paper; the text should explicitly state what was added relative to v1 (e.g., number of new images, new scenarios, range labels) rather than only saying 'greater variety.' Reference [7] and [8] are Roboflow Universe URLs and should note access dates consistently.","section":"References"}],"recommendation":"major_revision","confidential_remarks":"The paper is a dataset contribution rather than a methodological advance, which is appropriate for venues that accept dataset papers. The main risk is that the range-labeling claim, which is the paper's primary selling point, is entirely unsupported in the current preprint. This is fixable with a detailed sensor/measurement appendix and a validation study, but without that the paper should not be accepted. I would also encourage the editors to remind the authors that benchmark tables should report variance and that external evaluation sets should be clearly separated from training data."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Net take: this is a dataset release with a real niche contribution, and its main claim is currently under-supported. The new thing is LRDDv2, a second release of the authors' own LRDD set: 39,516 annotated images, 8,000+ with target range information, most with drones at or below 50 pixels in 1080p. That directly addresses a bottleneck in long-range UAV detection. The benchmark comparison is also legitimate: YOLOv8m trained on LRDDv2 or LRDDv2+Drone-vs-Bird beats training on Drone-vs-Bird alone when evaluated on Detfly and UAV-Detect. The use of external evaluation sets means no circularity problem.\n\nCredit where due: the range annotation attribute is not in Drone-vs-Bird, Anti-UAV, or UGV2, and the direction of the benchmark results is encouraging. The dataset URL pointing to a live project page is a plus.\n\nSoft spots, in proportion: the biggest one is exactly what the stress test flags. There is no description anywhere in the visible text of how the range labels were obtained, no measurement geometry, no GPS/RTK vs laser rangefinder vs structure-from-motion, and no validation or error analysis. Since the paper's headline feature and its stated future direction (range estimation) rest on those labels, that is a load-bearing omission. It does not sink the detection contribution, but it does mean the sentence \"making it possible to develop algorithms for drone range estimation\" is not currently supported by the preprint. The rest is of a piece with many dataset papers: no annotation protocol, no data split details, no per-run variability, and one descriptive analysis of detection probability vs bounding-box area that over-reads area as range. Those are ordinary weaknesses, not fatal ones.\n\nWho should read it: people working on UAV detection, small-object detection, and drone range estimation. The dataset likely is useful even if the range labels turn out to be noisy; the detection benchmark alone is worth knowing.\n\nI would not desk-reject this. Send it to review with a clear request for range-label methodology and validation, and for a public description of annotation and split. If the authors can document label accuracy, this is a solid contribution to a niche with few public resources.","headline":"Worth a serious look for the range-labeled long-range drone data, but the paper has to earn the range claim by documenting how the distances were measured and validated.","tokens_in":4957,"tokens_out":2995,"would_cite":false,"duration_ms":36353,"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":"LRDDv2 supplies 39,516 annotated drone images, with range labels on over 8,000, and models trained on it beat a standard drone dataset on two benchmarks.","keywords":["drone detection","long-range detection","dataset","range estimation","small object detection","UAV","YOLOv8","annotated images"],"falsifier":"Take a sample of the 8,000+ range-labeled images and compare each labeled distance against independent ground truth, such as a GPS flight log or laser rangefinder measurement taken at the same moment; if errors are large or systematic, the range-estimation claim collapses. A simpler check is whether labeled distance correlates monotonically with pixel size across the sample.","tokens_in":4080,"feed_emoji":"🛸","tokens_out":2905,"duration_ms":34387,"temperature":0.7,"pith_summary":"The paper introduces LRDDv2, a dataset of 39,516 annotated drone images built for long-range detection, where most drones occupy 50 or fewer pixels in 1080p resolution. Its standout feature is target range information for over 8,000 images, which the authors say makes range estimation possible for the first time in a drone detection dataset. Benchmarking with YOLOv8 shows that training on LRDDv2 alone outperforms training on the Drone-vs-Bird dataset when evaluated on the Detfly and UAV-Detect benchmarks, and that combining both datasets gives the best scores. The paper argues that this demonstrates the value of diverse, long-range-specific training data for real-world drone detection and safety.","feed_headline":"New drone dataset carries range info on 8,000+ images","feed_subtitle":"Models trained on LRDDv2 beat Drone-vs-Bird training on two long-range detection benchmarks.","key_machinery":"The central object is the dataset itself, specifically its composition: 39,516 annotated images with a deliberate majority of drones at 50 pixels or fewer in 1080p resolution, and a subset of over 8,000 images carrying explicit range labels. The paper uses bounding box area as a proxy for range, and its evaluation machinery is a fixed YOLOv8 model trained on different data sources and tested on the Detfly and UAV-Detect benchmarks; the range labels are the novel component that would enable distance-estimation algorithms.","core_discovery":"The central claim is that LRDDv2 improves long-range drone detection by providing a larger, more diverse set of small, distant drone images than existing datasets, and that this translates to measurable gains on independent benchmarks. Concretely, YOLOv8 trained on LRDDv2 reaches mAP@50 of 0.458 on Detfly versus 0.376 for Drone-vs-Bird training, and 0.562 versus 0.510 on UAV-Detect; combining Drone-vs-Bird with LRDDv2 yields 0.463 and 0.644 respectively. The paper also shows that detection probability falls as bounding box area shrinks, and that LRDDv2-trained models decay more gracefully at small sizes, supporting the claim that long-range-tailored data is a key ingredient for detecting distant UAVs.","pith_inferences":["If the range labels are accurate, LRDDv2 could be used to calibrate detection confidence against physical distance, letting safety systems decide when a drone is close enough to act on.","A natural validation step, not performed in the paper, would be to check whether labeled distance correlates monotonically with bounding box size; strong correlation would support the labels, while weak correlation would signal inconsistency.","The paper suggests patch-based training but does not test it; a follow-up could measure whether LRDDv2's small-object cases benefit even more from tiling than current benchmarks suggest.","The dataset's real utility for range estimation will only be clear once a model trained on its 8,000 labels is evaluated on held-out distances; that evaluation is absent from this release."],"forward_implications":["Researchers can train long-range drone detectors without collecting their own hard-to-find small-object imagery, since LRDDv2 supplies a large, varied set of such cases.","The 8,000+ range-labeled images open a route to supervised monocular distance estimation, a capability the paper identifies for air traffic control, collision avoidance, and surveillance.","Combining LRDDv2 with existing datasets like Drone-vs-Bird is a stronger recipe than either alone, suggesting that complementary data sources are valuable rather than redundant.","The detection-probability-versus-bounding-box analysis gives a concrete baseline for how far current YOLO-style detectors can see, and where performance degrades.","LRDDv2 can serve as a benchmark for future long-range detection datasets and for methods such as patch-based inference that are suggested for small-object detection."],"supporting_citations":[{"why":"Supplies the Drone-vs-Bird training dataset used as the baseline in the benchmarking experiments.","marker":"[6]"},{"why":"Supplies the UAV-Detect evaluation dataset used to test models trained on LRDDv2 and Drone-vs-Bird.","marker":"[7]"},{"why":"Supplies the Detfly evaluation dataset used as the second independent benchmark.","marker":"[8]"},{"why":"Describes LRDDv1, the predecessor dataset that LRDDv2 extends with more images and range information.","marker":"[14]"},{"why":"Describes the slicing-aided inference method the paper recommends for future small-object detection work, though it is not part of the main experiments.","marker":"[3]"}],"fun_headline_variants":["Range-aware drone dataset improves long-range UAV detection","LRDDv2: 39k images with range info for distant drone detection","New range-labeled drone dataset improves long-range detection","Range info in LRDDv2 yields better distant drone detection"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The distance values attached to the 8,000+ images are stated but their measurement or verification is not described; if these labels are inaccurate or inconsistent, the dataset's main novelty, enabling range estimation, would be compromised.","fun_headline_variants_meta":{"raw":{"variants":["Range-aware drone dataset improves long-range UAV detection","LRDDv2: 39k images with range info for distant drone detection","New range-labeled drone dataset improves long-range detection","Range info in LRDDv2 yields better distant drone detection"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000977,"raw_usage":{"total_tokens":4162,"prompt_tokens":970,"completion_tokens":3192,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":586,"completion_tokens_details":{"reasoning_tokens":3123}},"tokens_in":586,"tokens_out":3192,"duration_ms":26726,"temperature":1.0,"reasoning_tokens":3123,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-06T04:29:18.620051+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Take a sample of the 8,000+ range-labeled images and compare each labeled distance against independent ground truth, such as a GPS flight log or laser rangefinder measurement taken at the same moment; if errors are large or systematic, the range-estimation claim collapses. A simpler check is whether labeled distance correlates monotonically with pixel size across the sample.","supporting_citations":[{"cited_title":"bird detection: Deep learning algorithms and results from a grand challenge","cited_arxiv_id":null,"evidence_quote":"Supplies the Drone-vs-Bird training dataset used as the baseline in the benchmarking experiments."},{"cited_title":"https://universe.roboflow.com/get/uav-detect-pfiqs (jan 2023), https://universe.roboflow.com/get/uav-detect-pfiqs, visited on 2023-06-25","cited_arxiv_id":null,"evidence_quote":"Supplies the UAV-Detect evaluation dataset used to test models trained on LRDDv2 and Drone-vs-Bird."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supplies the Detfly evaluation dataset used as the second independent benchmark."},{"cited_title":"In: 2024 IEEE International Conference on Consumer Electronics (ICCE)","cited_arxiv_id":null,"evidence_quote":"Describes LRDDv1, the predecessor dataset that LRDDv2 extends with more images and range information."}],"review_version":1}