Pith. sign in

REVIEW 5 major objections 7 minor 1 cited by

Drone Detection and Tracking with YOLO and a Rule-based Method

T0 review · 5 major / 7 minor · reviewed 2026-08-08 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

Pith's one-line read Adding a cross-correlation trace to frame-by-frame YOLOv7 detection recovers 2 to 10 percent of drones that would otherwise be missed in video.

desk verdict A useful dataset extension and honest detection comparison, but the tracking claim is asserted, not measured. read the letter →

arxiv 2502.05292 v1 pith:UCHJR5O3 submitted 2025-02-07 cs.CV cs.AIcs.LG

Purbaditya Bhattacharya , Patrick Nowak This is my paper · ORCID
classification cs.CVcs.AIcs.LG
keywords dronedetectionobjectYOLOv7cross-correlationtrackinginfraredimagerycolordatasetextensionattentionmodules
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

This paper argues that a per-frame detector such as YOLOv7 can be made more reliable in drone videos by adding a cheap rule-based tracking layer: when confidence in a new detection dips, the layer uses cross-correlation between the previous frame's bounding-box neighborhood and the current frame to predict where the drone is and keeps the detection alive. The central claim, stated in Section V-A, is that this layer detects 2-10% more drones per video than YOLOv7 alone, with the gain depending on thresholds and video content. The paper also extends a published infrared drone dataset with annotated color images and a multi-channel set, and benchmarks YOLOv7 variants with attention modules. The practical point is that a detection drop caused by a sudden confidence dip can be repaired without retraining the network. If the claim holds, live drone surveillance can gain a simple post-processing improvement on top of the detector.

What carries the argument

The load-bearing mechanism is a two-threshold confidence rule combined with patch cross-correlation. The high threshold $t_h$ and low threshold $t_l$ divide detections into trusted detections, uncertain candidates, and lost objects. For an uncertain candidate, the previous frame's trusted box plus a slightly larger surrounding region is matched against the current frame by cross-correlation, which measures the best translational alignment of the two image patches. The peak location gives an estimated shift, and the previous box is moved by that shift to form a predicted box. This uses temporal continuity as a spatial prior that compensates for a one-frame drop in the detector's confidence. All threshold values are tunable, and the paper states the amount of improvement depends on their choice.

What would settle it

Take a labeled drone video with fast lateral motion, occlusion, or abrupt background changes and run YOLOv7 with and without the rule layer at fixed thresholds. If the rule layer does not produce more true-positive frames than YOLOv7 alone, or if its accepted boxes have low intersection over union with manual ground-truth boxes, the claimed 2-10% improvement fails. The paper's own caveat predicts failure whenever confidence stays below the low threshold for several consecutive frames, so a sustained low-confidence sequence is the simplest test.

Watch

Extended reading notes

Core claim

The central claim is that a rule-based, cross-correlation tracker layered on top of YOLOv7 recovers drones that YOLOv7 misses on individual frames. The method sets a high threshold and a low threshold. Boxes above the high threshold are trusted as valid detections and become references. If a current-frame box has confidence between the two thresholds, the method crops the region around the corresponding previous-frame box, cross-correlates it with the new frame, and shifts the previous box by the estimated translation. If the shifted box overlaps the current detection strongly (high intersection over union), the current detection is accepted with the previous high confidence; if the overlap is weak but the cross-correlation peak is strong, the predicted box is accepted instead. The paper reports this recovered an additional 2-10% of drones compared with YOLOv7 alone. It also reports that tracking stops when confidence falls below the low threshold, so the recovery applies to short dips, not sustained loss.

Load-bearing premise

The rule layer assumes the drone and its surrounding image patch move approximately as a plain shift from one frame to the next, and that high-confidence detections recur often enough to keep the tracker anchored.

Editorial extensions

If this is right

  • With the rule layer enabled, a YOLOv7-based system should detect 2-10% more drones per video than the bare detector, with the gain depending on the chosen thresholds and the video content.
  • The improvement comes without retraining: the same YOLOv7 weights and per-frame architecture are used, with only the post-processing rule layer added.
  • Because the rule layer relies on a single previous frame, tracking will stop whenever confidence falls below the low threshold, so the gain is limited to short confidence dips rather than sustained occlusions.
  • Among the tested YOLOv7 variants, adding a channel-attention module gives the best precision-to-speed trade-off, while a Swin transformer module costs substantially more inference time without better precision; the paper chooses channel attention for subsequent models.
  • The extended dataset now contains annotated color images plus a multi-channel infrared/color set, intended as a training resource, with the caveat that the multi-channel set was generated from a pre-trained YOLO model and should not be evaluated by retraining the same model.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

  • We infer the same two-threshold cross-correlation layer could be attached to any per-frame detector, not just YOLOv7, because it only consumes bounding boxes, confidence scores, and image patches.
  • We infer that using two or more previous frames, which the paper notes as a possible extension, would turn the shift estimate into a velocity estimate and should recover drones during longer occlusions and faster motion than the single-frame version.
  • We infer that the 2-10% range is scene-dependent enough that deployment would need scene-specific threshold calibration; in cluttered or rapidly changing backgrounds the translational-match assumption will be the limiting factor.
  • We infer that a properly synchronized multi-sensor camera system could make the multi-channel dataset usable for training, and the combined spectra would likely reduce the infrared false positives the paper reports.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Referee Report

5 major / 7 minor

Summary. This paper extends an existing infrared drone dataset with color and multichannel recordings, reports training and validation mAP for YOLOv7 and several attention/transformer variants, and proposes a rule-based cross-correlation tracker intended to reduce frame-to-frame detection drops. The central quantitative claim is that adding the tracker to YOLOv7 detects 2 to 10% more drones depending on the video, but this claim is supported only by a single illustrative frame pair. The multichannel dataset is generated from a pre-trained YOLO model and is not used in any training or evaluation experiment.

Significance. If the tracking claim were properly measured, the contribution would be modest but potentially useful: a lightweight cross-correlation layer on an off-the-shelf detector could recover some false-negative frames. The dataset documentation is informative, with per-camera statistics, annotation heatmaps, and clear mAP/latency tables for the detection models. The authors are also transparent about important limitations: they state that the multichannel data should not be evaluated by retraining the same model family, and that tracking stops when confidence falls below the low threshold. However, the paper's principal result is currently asserted rather than demonstrated, so the significance of the tracking contribution cannot be assessed from the manuscript as written.

major comments (5)
  1. [V-A] In the paragraph beginning "The proposed method running on top of YOLOv7...", the claim that the method detects 2 to 10% more drones is asserted without a measurement protocol. The text gives no definition of the detection-rate metric, no per-video table comparing YOLOv7 alone with YOLOv7 plus the rule layer, no values for conf_h, conf_l, the IOU threshold, or the cross-correlation threshold, and only one illustrative frame pair in Figure 9. The cross-correlation step is described only verbally, so the method cannot be reproduced. This is the paper's principal result and must be replaced by a complete evaluation with stated thresholds and per-video results.
  2. [V-A (rule bullets)] In the rule bullet list, a low-confidence box is promoted to valid with the previous high confidence score when the IOU with the predicted box is above threshold, and in the next bullet a predicted box is accepted when the cross-correlation value is high. Both acceptance paths can count boxes that are not actual drone detections as valid, so "more drones detected" could reflect an increase in false positives or propagated background instead of recovered true detections. The revision should report precision and recall on annotated frames at fixed thresholds and should separate recovered true positives from added false positives.
  3. [V-A, final paragraph] The method explicitly discontinues tracking when confidence falls below conf_l, and the authors acknowledge that many frames remain unsuccessful because of very low confidence. The claim that detection drops are reduced therefore applies only to a confidence corridor whose width is not quantified. The evaluation should run on complete test videos, report the fraction of frames for which the tracker is active, and compare the rule layer with at least one standard tracker or temporal baseline to show that the improvement is not simply an artifact of restricting the problem to easy frames.
  4. [III-B] The color annotation policy states that drones that are difficult to perceive are not annotated at all, while the validation split is drawn from these color images. Under this policy, a true detection of an unannotated difficult drone would be counted as a false positive, and a missed difficult drone would not be scored as a false negative, which biases the mAP numbers in Tables III and IV. The authors should annotate difficult instances with the "difficult" flag, which PASCAL VOC evaluation ignores, or explicitly describe how unannotated difficult examples were handled in the evaluation.
  5. [III-B, IV, VI] The multichannel dataset is a stated contribution, but no experiment uses it. It was generated with a pre-trained YOLO model, and the conclusion correctly notes that retraining the same model family on it would be circular; nevertheless, without an evaluation with an independent detector or a manually annotated multichannel test subset, the dataset's value remains untested. In addition, Section III-B says image pairs are discarded when registration yields poor results and asserts without evidence that the impact of background misalignment "should be less" for a single-class task. The revision should quantify this selection bias or test its effect.
minor comments (7)
  1. [Table I] Table I is captioned as the "final dataset used for training", but it includes the Multichannel row, which Section IV says is not used for training; please reword the caption or split the table.
  2. [Tables I-II] The infrared training and validation counts in Table II (44,300 + 12,483 = 56,783) do not reconcile with the 63,821 infrared images in Table I. The text mentions that some blurry or tiny-drone images were discarded, but the number of discarded images should be stated explicitly so the tables are consistent.
  3. [V] The sentence in Section V that "learning rates and the weight decay values... are adjusted" does not report the adjusted values; these hyperparameters should be listed for reproducibility, along with any other non-default training settings.
  4. [Overall] The abbreviation "MAP" is used where "mAP" is standard; please use consistent capitalization and define the metric once before the tables.
  5. [References] Reference [18] lists "An Image is Worth 16x16 Words" as ICLR 2022, but the paper appeared at ICLR 2021, and the page range appears to be copied from a different paper; please correct it.
  6. [Data availability] The manuscript does not state whether the extended dataset is publicly available; please add an availability statement or URL, since the paper's contribution is partly a dataset.
  7. [Figure 9] Figure 9 would be easier to interpret if the frames were numbered and the YOLOv7 confidence scores were shown for both the original detections and the detections recovered by the rule-based method.

Circularity Check

0 steps flagged · score 0.0 of 10

No circular derivation found; the tracking improvement is an unquantified empirical claim, not a result forced by the paper's definitions or self-citations.

full rationale

The paper's contributions are a dataset extension and a rule-based cross-correlation tracker layered on YOLOv7. The central claim that the tracker 'successfully detects more drones' is an empirical assertion about downstream detection continuity; it is not derived from the same equations that define the tracker, and no parameter is fitted to the claimed 2-10% improvement. The tracker's thresholds (conf_h, conf_l, IOU, cross-correlation) are free design choices, not fitted values that trivially produce the improvement metric. Self-citations to prior dataset work [1] provide provenance and annotation procedure, and the conclusion explicitly states that the multichannel dataset, generated by a pre-trained YOLO model, 'should not be evaluated by retraining with the same model,' which is an acknowledged limitation rather than a hidden circular evaluation. The lack of quantitative evaluation of the 2-10% claim is a correctness and reproducibility concern, not a circularity concern. Because no load-bearing step reduces to its own inputs by construction or via self-citation, no circular step is identified.

Assumptions & free parameters 6 free parameters · 3 assumptions · 0 invented entities

The paper rests on standard deep learning transfer, a translational-motion assumption for the tracker, and an assumption that discarding poorly registered image pairs does not bias the dataset. Several tunable thresholds and hyperparameters are not reported, and no code or dataset link is provided.

free parameters (6)
  • conf_h (high confidence threshold) = not reported
    Defines which detections are valid in the rule-based tracker; the authors state thresholds can be experimented with but give no values, so the reported 2-10% improvement is threshold-dependent.
  • conf_l (low confidence threshold) = not reported
    Lower bound for candidate detections; tracking stops below this value, so it directly controls how many frames the tracker can recover.
  • IOU threshold for box validation = not reported
    Used to decide whether a cross-correlation predicted box matches the detector's new box.
  • Cross-correlation threshold = not reported
    If the IOU threshold is not met, the predicted box is accepted only when correlation exceeds this threshold.
  • Learning rate = not reported
    The paper says learning rates and weight decay were adjusted but does not state the values, preventing replication of training.
  • Weight decay = not reported
    Same as learning rate; adjusted but unreported, so the exact training configuration is unknown.
assumptions (3)
  • domain assumption Consecutive-frame drone motion and local background change are well approximated by a translation recoverable by cross-correlation.
    The tracker in Section V-A predicts bounding boxes from the estimated shift between corresponding local regions; the paper concedes it fails on very low confidence frames.
  • domain assumption COCO-pretrained YOLOv7 weights transfer usefully to drone detection.
    Section V says the model is pretrained on the COCO dataset and the pretrained weights are used to initialize during training.
  • ad hoc to paper Discarding image pairs with poor registration does not bias the multichannel dataset or the conclusions.
    Section III-B says image pairs are discarded when a registration yields poor results, and the impact of such a problem should be less; this discarding is not analyzed for selection bias.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Drone Detection and Tracking with YOLO and a Rule-based Method." pith.science (2026). https://pith.science/paper/UCHJR5O3

@misc{pith2026250205292,
  author       = {Pith},
  title        = {Pith review of: Drone Detection and Tracking with YOLO and a Rule-based Method},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/UCHJR5O3}},
  note         = {Machine review of arXiv:2502.05292}
}
read the original abstract

Drones or unmanned aerial vehicles are traditionally used for military missions, warfare, and espionage. However, the usage of drones has significantly increased due to multiple industrial applications involving security and inspection, transportation, research purposes, and recreational drone flying. Such an increased volume of drone activity in public spaces requires regulatory actions for purposes of privacy protection and safety. Hence, detection of illegal drone activities such as boundary encroachment becomes a necessity. Such detection tasks are usually automated and performed by deep learning models which are trained on annotated image datasets. This paper builds on a previous work and extends an already published open source dataset. A description and analysis of the entire dataset is provided. The dataset is used to train the YOLOv7 deep learning model and some of its minor variants and the results are provided. Since the detection models are based on a single image input, a simple cross-correlation based tracker is used to reduce detection drops and improve tracking performance in videos. Finally, the entire drone detection system is summarized.

Figures

Figures reproduced from arXiv: 2502.05292 by the authors.

Figure 1
Figure 1. Overview of the drone detection system. videos are recorded at the harbour location with additional backgrounds and the corresponding frames are annotated and added to the dataset. With the incremental improvement in deep learning based methods, they are primarily used for multiple object detection tasks [2]–[5]. More commercial multi-sensor or vision systems now use a trained and properly scaled deep learning model… view at source ↗
Figure 3
Figure 3. Used cameras: FLIR Scion OTM366 (left), InfraTec [PITH_FULL_IMAGE:figures/full_fig_p002_3.png] view at source ↗
Figure 4
Figure 4. Exemplary images of the dataset recorded with FLIR Scion OTM366 at (a) harbour, with InfraTec VarioCAM HD Z at [PITH_FULL_IMAGE:figures/full_fig_p003_4.png] view at source ↗
Figures from the paper (6 more)
Figure 5
Figure 5. Figure 5: Annotation heatmap of the drones recorded by the [PITH_FULL_IMAGE:figures/full_fig_p004_5.png]
Figure 6
Figure 6. Figure 6: Annotation heatmap of the drones recorded by the Sony [PITH_FULL_IMAGE:figures/full_fig_p004_6.png]
Figure 7
Figure 7. Figure 7: A simple illustration of the different attention modules. [PITH_FULL_IMAGE:figures/full_fig_p005_7.png]
Figure 8
Figure 8. Figure 8: Example of drone detection on some cropped sections of individual images from test data by the YOLOv7 model. [PITH_FULL_IMAGE:figures/full_fig_p007_8.png]
Figure 9
Figure 9. Figure 9: Consecutive frames from a test video having missing detections with YOLOv7 (top) and the same frames with successful [PITH_FULL_IMAGE:figures/full_fig_p008_9.png]
Figure 10
Figure 10. Figure 10: A simple test for live drone detection. be constructed with a real scenario and accurate measurement setup. VII. ACKNOWLEDGEMENT This work was part of the “Digital Sensor-2-Cloud Campus Platform” (DS2CCP) project, which is funded by the Federal Ministry of Defense und…

Discussion (0). Continue with ORCID to comment.

Forward citations

Cited by 1 Pith paper

Reviewed papers in the Pith corpus that reference this work. Sorted by Pith novelty score. Full citation record

  1. WAVE-DETR Multi-Modal Visible and Acoustic Real-Life Drone Detector

    cs.CV 2025-09 conditional novelty 6.0 of 10

    Fusing Wav2Vec2 audio features into a Deformable DETR detector improves small-drone detection mAP by 11-15% on a new synchronized audio-RGB dataset.

Reference graph

Works this paper leans on

20 extracted references · 17 canonical work pages · cited by 1 Pith paper

  1. [1]

    A Dataset of Infrared Images for Deep Learning based Drone Detection,

    P. Bhattacharya, P. Nowak, D. Ahlers, and U. Z ¨olzer, “A Dataset of Infrared Images for Deep Learning based Drone Detection,” in 2023 17th International Conference on Signal-Image Technology and Internet- Based Systems (SITIS) , Bangkok, Thailand, pp. 130–137, 2023

    work page 2023

  2. [2]

    Faster R-CNN: Towards Real- Time Object Detection with Region Proposal Networks,

    S. Ren, K. He, R. Girshick, and J. Sun, “Faster R-CNN: Towards Real- Time Object Detection with Region Proposal Networks,”, in Advances in Neural Information Processing Systems (NIPS) , vol. 28, 2015

    work page 2015

  3. [3]

    SSD: Single Shot MultiBox Detector,

    W. Liu et al., “SSD: Single Shot MultiBox Detector,” in European Conference on Computer Vision (ECCV) , pp. 21–37, 2016

    work page 2016

  4. [4]

    EfficientDet: Scalable and Efficient Object Detection,

    M. Tan, R. Pang, and Q. V . Le, “EfficientDet: Scalable and Efficient Object Detection,” in 2020 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) , pp. 10778–10787, 2020

    work page 2020

  5. [5]

    Shore Livecams: A Maritime Dataset for Deep Learning based Object Detection,

    D. Ahlers, P. Bhattacharya, P. Nowak, and U. Z ¨olzer, “Shore Livecams: A Maritime Dataset for Deep Learning based Object Detection,” in 2023 17th International Conference on Signal-Image Technology and Internet- Based Systems (SITIS) , Bangkok, Thailand, pp. 138-144, 2023

    work page 2023

  6. [6]

    CNN based Drone Detection in Infrared Images,

    P. Bhattacharya, P. Nowak, D. Ahlers, M. Holters, and U. Z ¨olzer, “CNN based Drone Detection in Infrared Images,” in dtec.bw-Beitr¨age der Helmut-Schmidt-Universit¨at / Universit ¨at der Bundeswehr Hamburg: F orschungsaktivit¨aten im Zentrum f ¨ur Digitalisierungs- und Technolo- gieforschung der Bundeswehr dtec.bw - Band 1 , pp. 141—146, 2022

    work page 2022

  7. [7]

    Real-Time Detection for Small UAVs: Combining YOLO and Multi-frame Motion Analysis

    J. Liu, L. Plotegher, E Roura, C. de Souza Junior, and S He, “Real-Time Detection for Small UA Vs: Combining YOLO and Multi-frame Motion Analysis”, in ArXiv, arXiv:2411.02582v1, 2024

    work page Pith review arXiv 2024

  8. [8]

    GStreamer: open source multimedia framework

    GStreamer, “GStreamer: open source multimedia framework”, Available: https://gstreamer.freedesktop.org/ [Visited on 01/28/2025]

    work page 2025

Show all 20 references
  1. [9]

    Docker: lightweight linux containers for consistent devel- opment and deployment

    D. Merkel, “Docker: lightweight linux containers for consistent devel- opment and deployment”, in Linux journal , vol. 2014 (239), p.2, 2014

    2014

  2. [10]

    LabelImg,

    Tzutalin, “LabelImg,” Free Software: MIT License, 2015, [On- line], Available: https://github.com/HumanSignal/labelImg [Visited on 01/28/2025]

    2015

  3. [11]

    The Pascal Visual Object Classes (VOC) Challenge,

    M. Everingham, L. van Gool, C. K. I. Williams, J. Winn, and A. Zisserman, “The Pascal Visual Object Classes (VOC) Challenge,” in International Journal of Computer Vision , vol. 88, pp. 303—338, 2010

    2010

  4. [12]

    ultralytics/yolov5: v3.1 - Bug Fixes and Performance Improvements,

    G. Jocker et al., “ultralytics/yolov5: v3.1 - Bug Fixes and Performance Improvements,” Zenodo, October 2020, https://doi.org/10.5281/zenodo.4154370

    2020 doi

  5. [13]

    YOLOv6: A Single-Stage Object Detection Framework for Industrial Applications,

    C. Li et al., “YOLOv6: A Single-Stage Object Detection Framework for Industrial Applications,” in ArXiv, 2209.02976, 2022

    2022 arXiv

  6. [14]

    YOLOv7: Trainable Bag-of-Freebies sets New State-of-the-Art for Real-Time Object Detec- tors,

    C.-Y . Wang, A. Bochkovskiy, and H.-Y . M. Liao, “YOLOv7: Trainable Bag-of-Freebies sets New State-of-the-Art for Real-Time Object Detec- tors,” in ArXiv, 2207.02696, 2022

    2022 arXiv

  7. [15]

    Deep Learning for Image Enhancement,

    P. Bhattacharya, “Deep Learning for Image Enhancement,” Helmut- Schmidt-Universit¨at / Universit ¨at der Bundeswehr Hamburg, 2022, doi: 10.24405/14267

    2022 doi

  8. [16]

    CBAM: Convolutional Block Attention Module,

    S. Woo, J. Park, J-Y . Lee, I. S. Kweon, “CBAM: Convolutional Block Attention Module,” in Proceedings of the European Conference on Computer Vision (ECCV) , September, 2018

    2018

  9. [17]

    Attention is All you Need,

    A. Vaswani et al., “Attention is All you Need,” in Proceedings of the 31st International Conference on Neural Information Processing Systems (NIPS), pp. 6000–6010, 2017

    2017

  10. [18]

    An Image is Worth 16x16 Words: Transformers for Image Recognition at Scale,

    A. Dosovitskiy et al., “An Image is Worth 16x16 Words: Transformers for Image Recognition at Scale,” in 9th International Conference on Learning Representations (ICLR) , Austria, pp. 711—727, 2022

    2022

  11. [19]

    A Simple Single-Scale Vision Transformer for Object Detection and Instance Segmentation,

    W. Chen et al., “A Simple Single-Scale Vision Transformer for Object Detection and Instance Segmentation,” in Proceedings of the European Conference on Computer Vision (ECCV), Tel-Aviv, Israel, pp. 711—727, 2022

    2022

  12. [20]

    Swin Transformer: Hierarchical Vision Transformer using Shifted Windows,

    Z. Liu et al., “Swin Transformer: Hierarchical Vision Transformer using Shifted Windows,” in 2021 IEEE/CVF International Conference on Computer Vision (ICCV), Montreal, QC, Canada, pp. 9992-10002, 2021

    2021

Pith tools

Reviewed August 8, 2026 · model on record in the stance chip above.