Pith. sign in

REVIEW 5 major objections 7 minor 3 cited by

Separating Drone Point Clouds From Complex Backgrounds by Cluster Filter -- Technical Report for CVPR 2024 UG2 Challenge

T0 review · 5 major / 7 minor · reviewed 2026-08-11 · deepseek-v4-flash

Pith's one-line read The paper claims a cluster filter can isolate a drone's point cloud from noisy lidar background by scoring each candidate on density ratio and voxel displacement, then picking the highest score as the trajectory.

desk verdict Unsupervised drone-trajectory extraction with a plausible density/IoU scoring rule and an external 4th-place challenge finish, but the report is missing nearly all quantitative support. read the letter →

arxiv 2412.16947 v1 pith:E7CS3CB6 submitted 2024-12-22 cs.CV

classification cs.CV
keywords UAVdetectionlidarpointcloudunsupervisedclusteringDBSCANspatiotemporalvoxelanalysisdronetrajectoryestimationdenoisinganti-dronesystems
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

The paper tries to establish that a small UAV's point cloud can be pulled out of noisy, multi-lidar backgrounds without any supervised training. Its method stacks all scans in a sequence, clusters the stacked cloud with DBSCAN, then looks at each cluster again in short time windows and computes two numbers: how similar the cluster's local density is to its global density, and how much its occupied voxels shift between windows. A moving drone is expected to keep its density ratio high while its voxels move, whereas static clutter accumulates density and keeps its voxels still, so the cluster maximizing a weighted sum of these signals should be the drone trajectory. The authors report the approach finished 4th in the 2024 UG2+ challenge on the MMAUD dataset, with a 99.15% sequence detection accuracy, supporting the claim that a simple interpretable pipeline can compete with deep detectors.

What carries the argument

The mechanism is a global-local cluster filter: DBSCAN, a density-based clustering algorithm, runs on the superimposed global cloud to define candidate objects, and each candidate is re-examined inside short sliding time windows of the original sequence. The two load-bearing quantities are the relative density ratio $R_k^{\mathrm{frame}} = \rho_k^{\mathrm{local}} / \rho_k^{\mathrm{global}}$ and the voxel IoU $\mathrm{IoU}^{i,j}_k$ between local windows. They are combined as $\mathrm{Score}^k = \sum e^{R_k^{\mathrm{frame}}} + \lambda \sum \log(1/\mathrm{IoU}^{i,j}_k)$, a weighted sum that rewards a cluster whose local density stays comparable to its global density while its occupied voxels shift. Spline fitting over timestamp-ordered selected points turns the selected cluster into the 3D trajectory. This mechanism is what lets an unsupervised pipeline separate a moving small target from static clutter.

What would settle it

Take a lidar sequence that contains one drone and one moving distractor, such as a bird or vehicle, and compute the paper's Score for every DBSCAN cluster; if any non-drone cluster outscores the true drone cluster in even one time window, the highest-score-is-the-drone claim fails. The absence of per-sequence scores in the paper makes this check the natural way to test the claim.

Watch

Extended reading notes

Core claim

The central claim is that the drone trajectory is exactly the cluster with the highest combined score of density-ratio stability and voxel displacement. In the method, all point clouds in a sequence are superimposed and clustered by DBSCAN; each resulting global cluster is then intersected with short temporal windows of the original sequence. For every cluster, the paper computes the local density, the relative density ratio $R_k^{\mathrm{frame}} = \rho_k^{\mathrm{local}} / \rho_k^{\mathrm{global}}$, and the voxel Intersection-over-Union between windows. The scoring identity is $\mathrm{Score}^k = \sum e^{R_k^{\mathrm{frame}}} + \lambda \sum \log(1/\mathrm{IoU}^{i,j}_k)$, and the paper states that the target with the highest confidence is selected as the final target, that is, the drone trajectory. The selected cluster is then sorted by timestamp and fitted with a spline to produce the UAV's 3D coordinates at arbitrary time nodes. The paper presents this as an unsupervised alternative to deep detection, arguing that small drones are too sparse and unstable for traditional detectors but leave a distinctive motion-density trace in the point-cloud time series.

Load-bearing premise

The load-bearing premise is that the cluster with the highest density-and-voxel score is always the drone; if a bird, vehicle, or any other moving object ever has the right density and overlap profile, the scorer would pick it instead of the drone.

Editorial extensions

If this is right

  • Anti-drone detection from lidar can run without labeled training data, so it transfers to new drones and new sites without re-annotation.
  • The pipeline outputs the UAV's 3D trajectory directly, so downstream tracking or interception systems can use it without a separate data-association step.
  • Because the computation is clustering plus arithmetic, it can be deployed on edge devices rather than requiring a GPU or a large deep network.
  • The reported 99.15% sequence detection accuracy on the challenge benchmark indicates the unsupervised scorer misses the drone during only a small fraction of the sequence time.

Reading between the lines

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

  • The method's decision rule assumes one drone per sequence; scenes with two drones or several movers would require a top-k extension or a second discriminative stage, which the paper does not discuss.
  • The same cluster-scoring signature should pick out any small moving object, such as a bird, ground vehicle, or debris, not only drones, because it never models what a drone looks like; that generality could be tested directly on diverse moving-target lidar sequences.
  • The paper reports no per-sequence scores or sensitivity of the weight $\lambda$; without them, a reader cannot tell how close the runner-up clusters are or how robust the highest-score rule is across the 50 sequences.
  • Combining the cluster score with audio or radar modalities, which the MMAUD dataset also provides, could disambiguate cases where a bird and a drone have similar motion-density profiles; that fusion is a natural next step not explored by the paper.
Share X Bluesky LinkedIn Reddit HN

Signed reviews

No signed human review yet.

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. The paper proposes an unsupervised lidar point cloud pipeline for small-drone detection and trajectory reconstruction. It denoises Livox Avia data, performs global and local DBSCAN clustering over spatio-temporal point cloud sequences, computes per-cluster density ratios and voxel IoU between frames, and combines them into a confidence score whose maximum is declared to be the drone trajectory. The selected point cloud is then fitted with a cubic spline and interpolated to timestamps to produce estimated 3D positions. The authors report a 4th-place finish in the CVPR 2024 UG2+ Challenge on the MMAUD dataset and claim the method is simple, fast, and interpretable.

Significance. If the described claim were fully validated, the paper would offer a useful unsupervised alternative to supervised detectors for a difficult small-object lidar task, and the interpretable scoring mechanism would be a noteworthy contribution. The method's reliance on only two hand-picked cues (density stability and inter-frame voxel overlap) is elegant and could be applicable beyond drones. However, the evidence presented is not currently sufficient: the quantitative table is corrupted, the scoring equations contain a cluster-index summation error, the selection rule is unvalidated, and all hyperparameters are unreported. The external challenge result suggests a working system existed, but it does not by itself establish that the described scoring rule, rather than unstated preprocessing or calibration, is what selects the correct trajectory.

major comments (5)
  1. [Section III-C, Eq. (6)-(8)] The definitions Score_IoU^k = sum_{k=1}^{n} log(1/IoU^{i,j}_k) and Score_dens^k = sum_{k=1}^{n} exp(R_k^frame) sum over the cluster index k inside a per-cluster score. As written, the index k is summed out, so the score is not cluster-specific and the sentence "the category with the highest score" (Section III-C) is ill-defined. The summation should be over frames or time windows, and the score must be indexed by the cluster k.
  2. [Section III-C, Eq. (6)] The term log(1/IoU) is singular and unbounded as IoU approaches zero. A cluster whose voxel sets are disjoint in adjacent frames receives an arbitrarily large IoU score, so this term can dominate and make the ranking unstable. The paper gives no epsilon offset, clipping rule, or averaging over frames to make the score finite and comparable across clusters, which is load-bearing for the claimed selection rule.
  3. [Table I] The comparison table is corrupted: every method has MSE value '1', several rows have no method name, the backbone column contains stray values like '20', and the proposed method's SDA of 99.15% is given without standard deviation, number of sequences, or per-sequence results. This table provides no quantitative support for the central trajectory-accuracy claim, and no MSE numbers or error bars appear elsewhere in the paper.
  4. [Section III-B/III-C, central selection claim] The paper asserts without supporting evidence that the cluster with the maximum Score is always the drone trajectory. The two scoring cues (stable local density and low inter-frame voxel IoU) are also exhibited by other small moving objects such as birds or debris, and the paper provides no per-sequence score distribution, no failure analysis, and no sensitivity study for the weighting lambda. The external 4th-place result cannot substitute for validating the specific selection rule described in the manuscript.
  5. [Sections III-A and III-B, hyperparameters] The method depends on several hand-chosen hyperparameters that are never reported: DBSCAN epsilon and MinPts, the local window length 'frames', the voxel size used for V_global and V_local, the noise-density threshold for denoising, and the scoring weight lambda. Without these values the method is not reproducible, and the absence of a sensitivity analysis leaves open whether the reported challenge result relies on a narrow parameter choice.
minor comments (7)
  1. [Throughout] The manuscript repeatedly uses 'UA V' and 'MA V' with spurious spaces; the text should use 'UAV' and 'MAV' consistently.
  2. [Section III-D] The spline basis function B_i(u) includes an undefined parameter h and a four-element polynomial list; it does not match standard cubic B-spline notation and should be rewritten or given a proper citation.
  3. [Section III-C] The text says 'as shown in Figure ()', but the figure pointer is left incomplete; Figure 4 is referenced later but its actual content is not embedded in the manuscript.
  4. [Section III-C] The text refers to 'Section 3.2' when describing where the scoring mechanism is detailed, but the scoring mechanism is in Section III-C.
  5. [Table I] Several rows in Table I have empty method names and stray '20' values in the backbone column; these entries should be completed or removed.
  6. [Contributions list, Section I] The fourth contribution claims 'ablation experiments', but no ablation experiments appear anywhere in the paper; this claim should be removed or the experiments should be added.
  7. [References] References [12], [18], and [19] are cited as ICASSP 2025 papers without page or DOI information and may not be publicly available yet; the authors should verify these entries.

Circularity Check

0 steps flagged · score 0.0 of 10

No circularity found: the scoring heuristic is falsifiable and externally benchmarked, and the only self-citation is not load-bearing.

full rationale

The derivation chain is not circular. The pipeline computes density and voxel statistics directly from the input point clouds: global density rho^k_global = Num^k_global / V^k_global, local density rho^{k,frame}_local, IoU^{i,j}_k, relative density R, and the combined score Scorek = Scorek_dens + lambda * Scorek_IoU; none of these quantities is fitted to the ground-truth drone trajectory and then renamed as a prediction. The decision "the target with the highest confidence is selected as the final target. That is, the drone trajectory" is an empirical selection heuristic: it is checked against MMAUD ground truth via MSE/SDA and by the external CVPR 2024 UG2+ Challenge, so the statement is falsifiable rather than true by definition. The only self-citation ([18], H. Liang et al., sharing two authors) appears in the related-work discussion of lidar noise and sparsity and does not supply the scoring rule or any uniqueness claim, so it is not load-bearing. The displayed Scorek formulas sum over the cluster index k inside a per-cluster score, which as written makes the expression ill-defined; this is a correctness/clarity defect, not a circular reduction. Unreported hyperparameters and the absence of per-sequence score distributions are reproducibility and validation concerns, but they do not make any prediction equivalent to its input.

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

The paper introduces no new physical entities; it relies on six hand-chosen or unstated free parameters and five assumptions grounded in standard clustering math plus domain-specific sensor behavior. The most consequential load-bearing item is the assumption that the highest-scoring cluster is the drone, which is asserted without validation.

free parameters (6)
  • DBSCAN epsilon (ε)
    Neighborhood radius for global and local clustering; value not reported in Section III-B, must be hand-tuned.
  • DBSCAN MinPts
    Minimum points to form a cluster; not reported; controls noise removal and cluster granularity.
  • Local time window length 'frames'
    Defines P_local in Section III-B; not specified, changes density and IoU statistics.
  • Voxel size for V_global and V_local
    Volume computation in Section III-B requires a voxel grid resolution that is not given.
  • Noise density threshold for denoising
    Section III-A removes the 'sparsest' density class; the cutoff is not quantified.
  • Scoring weight lambda (λ)
    Balances density and IoU scores in Score^k = Score_dens + lambda * Score_IoU (Section III-C); value not provided.
assumptions (6)
  • domain assumption Moving-object point cloud density is approximately constant over short windows while static background density increases.
    Stated in Section III-B as the basis for relative density ratio R^frame_k.
  • ad hoc to paper The cluster with the highest combined score is the drone trajectory.
    Section III-C: 'the target with the highest confidence is selected as the final target'; no proof, validation, or failure cases are given.
  • domain assumption Static background voxel overlap (IoU) is higher across adjacent windows than the moving drone's.
    Used to build Score_IoU in Section III-C; relies on unvalidated scene statistics.
  • standard math DBSCAN yields meaningful clusters on the fused multi-lidar point cloud.
    Section III-B chooses DBSCAN for arbitrary-shaped clusters; standard algorithm assumed effective on this data.
  • standard math Cubic B-spline interpolation over timestamp-ordered points recovers drone positions at arbitrary frames.
    Section III-D defines the spline; assumes smooth trajectory and correct control points.
  • domain assumption MMAUD ground truth and timestamps are accurate.
    Evaluation in Section IV depends on dataset ground truth; not verified.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Separating Drone Point Clouds From Complex Backgrounds by Cluster Filter -- Technical Report for CVPR 2024 UG2 Challenge." pith.science (2026). https://pith.science/paper/E7CS3CB6

@misc{pith2026241216947,
  author       = {Pith},
  title        = {Pith review of: Separating Drone Point Clouds From Complex Backgrounds by Cluster Filter -- Technical Report for CVPR 2024 UG2 Challenge},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/E7CS3CB6}},
  note         = {Machine review of arXiv:2412.16947}
}
read the original abstract

The increasing deployment of small drones as tools of conflict and disruption has amplified their threat, highlighting the urgent need for effective anti-drone measures. However, the compact size of most drones presents a significant challenge, as traditional supervised point cloud or image-based object detection methods often fail to identify such small objects effectively. This paper proposes a simple UAV detection method using an unsupervised pipeline. It uses spatial-temporal sequence processing to fuse multiple lidar datasets effectively, tracking and determining the position of UAVs, so as to detect and track UAVs in challenging environments. Our method performs front and rear background segmentation of point clouds through a global-local sequence clusterer and parses point cloud data from both the spatial-temporal density and spatial-temporal voxels of the point cloud. Furthermore, a scoring mechanism for point cloud moving targets is proposed, using time series detection to improve accuracy and efficiency. We used the MMAUD dataset, and our method achieved 4th place in the CVPR 2024 UG2+ Challenge, confirming the effectiveness of our method in practical applications.

Figures

Figures reproduced from arXiv: 2412.16947 by the authors.

Figure 1
Figure 1. Challenging examples of image and point cloud detection. In [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. We first superimpose all point clouds in the sequence to obtain the global point cloud, then separate the data from different lidars, and only [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗
Figure 3
Figure 3. Our proposed algorithm architecture. Given a continuous point cloud input sequence, we first classify it into global clustering and local clustering [PITH_FULL_IMAGE:figures/full_fig_p004_3.png] view at source ↗
Figures from the paper (1 more)
Figure 4
Figure 4. Figure 4: The green point cloud in Figure b is the UAV point cloud separated [PITH_FULL_IMAGE:figures/full_fig_p006_4.png]

Discussion (0). Continue with ORCID to comment.

Forward citations

Cited by 3 Pith papers

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

  1. TAME: Temporal Audio-based Mamba for Enhanced Drone Trajectory Estimation and Classification

    cs.SD 2024-12 conditional novelty 5.0 of 10

    TAME applies parallel Mamba state-space models to audio spectrograms and reports state-of-the-art drone trajectory estimation and classification on MMAUD, with unresolved evaluation concerns.

  2. Audio Array-Based 3D UAV Trajectory Estimation with LiDAR Pseudo-Labeling

    cs.RO 2024-12 conditional novelty 5.0 of 10

    An audio-only neural network predicts 3D UAV trajectories with 0.48 m average position error on MMAUD, trained with unsupervised LiDAR trajectory estimates as pseudo-labels.

  3. Unsupervised UAV 3D Trajectories Estimation with Sparse Point Clouds

    cs.CV 2024-12 conditional novelty 4.0 of 10

    An unsupervised LiDAR clustering and spline method estimates UAV 3D trajectories from sparse point clouds, reporting 1.35 m RMSE on the MMAUD v2/v3 benchmark.

Reference graph

Works this paper leans on

19 extracted references · 17 canonical work pages · cited by 3 Pith papers

  1. [18]

    Unsupervised uav 3d trajectories estimation with sparse point clouds,

    H. Liang, Y . Yang, J. Hu, J. Yang, F. Liu, and S. Yuan, “Unsupervised uav 3d trajectories estimation with sparse point clouds,” 2025 IEEE International Conference on Acoustics, Speech, and Signal Processing , 2025

  2. [1]

    Air-to-air visual detection of micro-uavs: An experimental evaluation of deep learning,

    Y . Zheng, Z. Chen, D. Lv, Z. Li, Z. Lan, and S. Zhao, “Air-to-air visual detection of micro-uavs: An experimental evaluation of deep learning,” IEEE Robotics and automation letters , vol. 6, no. 2, pp. 1020–1027, 2021

  3. [2]

    Unmanned aerial vehicle visual detection and tracking using deep neural networks: A performance benchmark,

    B. K. Isaac-Medina, M. Poyser, D. Organisciak, C. G. Willcocks, T. P. Breckon, and H. P. Shum, “Unmanned aerial vehicle visual detection and tracking using deep neural networks: A performance benchmark,” in Proceedings of the IEEE/CVF International Conference on Com- puter Vision, 2021, pp. 1223–1232

  4. [3]

    Real-time small drones detection based on pruned yolov4,

    H. Liu, K. Fan, Q. Ouyang, and N. Li, “Real-time small drones detection based on pruned yolov4,” Sensors, vol. 21, no. 10, p. 3374, 2021

  5. [4]

    A Comprehensive Approach for UAV Small Object Detection with Simulation-based Transfer Learning and Adaptive Fusion

    C. Rui, G. Youwei, Z. Huafei, and J. Hongyu, “A comprehensive approach for uav small object detection with simulation-based transfer learning and adaptive fusion,” arXiv preprint arXiv:2109.01800, 2021

  6. [5]

    Real- time and accurate drone detection in a video with a static background,

    U. Seidaliyeva, D. Akhmetov, L. Ilipbayeva, and E. T. Matson, “Real- time and accurate drone detection in a video with a static background,” Sensors, vol. 20, no. 14, p. 3856, 2020. 6

  7. [6]

    Small low-contrast target detection: Data-driven spatiotemporal feature fusion and implementa- tion,

    J. Xie, C. Gao, J. Wu, Z. Shi, and J. Chen, “Small low-contrast target detection: Data-driven spatiotemporal feature fusion and implementa- tion,” IEEE transactions on cybernetics , vol. 52, no. 11, pp. 11 847– 11 858, 2021

  8. [7]

    Adaptive switching spatial-temporal fusion detection for remote flying drones,

    J. Xie, J. Yu, J. Wu, Z. Shi, and J. Chen, “Adaptive switching spatial-temporal fusion detection for remote flying drones,” IEEE Transactions on V ehicular Technology, vol. 69, no. 7, pp. 6964–6976, 2020

Show all 19 references
  1. [8]

    Detection, localization, and tracking of multiple mavs with panoramic stereo camera networks,

    Y . Zheng, C. Zheng, X. Zhang, F. Chen, Z. Chen, and S. Zhao, “Detection, localization, and tracking of multiple mavs with panoramic stereo camera networks,” IEEE transactions on automation science and engineering , vol. 20, no. 2, pp. 1226–1243, 2022

  2. [9]

    Fast and robust uav to uav detection and tracking from video,

    J. Li, D. H. Ye, M. Kolsch, J. P. Wachs, and C. A. Bouman, “Fast and robust uav to uav detection and tracking from video,” IEEE Transactions on Emerging Topics in Computing , vol. 10, no. 3, pp. 1519–1531, 2021

  3. [10]

    Dogfight: Detecting drones from drones videos,

    M. W. Ashraf, W. Sultani, and M. Shah, “Dogfight: Detecting drones from drones videos,” in Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , 2021, pp. 7067–7076

  4. [11]

    Low in resolution, high in precision: Uav detection with super-resolution and motion information extraction,

    H. Wang, X. Wang, C. Zhou, W. Meng, and Z. Shi, “Low in resolution, high in precision: Uav detection with super-resolution and motion information extraction,” in ICASSP 2023-2023 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP) . IEEE, 2023, pp. 1–5

  5. [12]

    Audio array-based 3d uav trajectory estimation with lidar pseudo-labeling,

    A. Lei, T. Deng, H. Wang, J. Yang, and S. Yuan, “Audio array-based 3d uav trajectory estimation with lidar pseudo-labeling,” 2025 IEEE International Conference on Acoustics, Speech, and Signal Processing , 2025

  6. [13]

    Adaptive lidar scan frame integration: Tracking known mavs in 3d point clouds,

    L. Qingqing, Y . Xianjia, J. P. Queralta, and T. Westerlund, “Adaptive lidar scan frame integration: Tracking known mavs in 3d point clouds,” in 2021 20th International Conference on Advanced Robotics (ICAR) . IEEE, 2021, pp. 1079–1086

  7. [14]

    Drone detection using sparse lidar mea- surements,

    S. Dogru and L. Marques, “Drone detection using sparse lidar mea- surements,” IEEE Robotics and Automation Letters , vol. 7, no. 2, pp. 3062–3069, 2022

  8. [15]

    Detection and tracking of small objects in sparse 3d laser range data,

    J. Razlaw, J. Quenzel, and S. Behnke, “Detection and tracking of small objects in sparse 3d laser range data,” in 2019 International Conference on Robotics and Automation (ICRA) . IEEE, 2019, pp. 2967–2973

  9. [16]

    Study on target detection and tracking method of uav based on lidar,

    H. Wang, Y . Peng, L. Liu, and J. Liang, “Study on target detection and tracking method of uav based on lidar,” in 2021 Global Reliability and Prognostics and Health Management (PHM-Nanjing) . IEEE, 2021, pp. 1–6

  10. [17]

    Uav tracking with lidar as a camera sensor in gnss-denied environ- ments,

    H. Sier, X. Yu, I. Catalano, J. P. Queralta, Z. Zou, and T. Westerlund, “Uav tracking with lidar as a camera sensor in gnss-denied environ- ments,” in 2023 International Conference on Localization and GNSS (ICL-GNSS). IEEE, 2023, pp. 1–7

  11. [19]

    Tame: Temporal audio-based mamba for enhanced drone trajectory estimation and classification,

    Z. Xiao, H. Hu, G. Xu, and J. He, “Tame: Temporal audio-based mamba for enhanced drone trajectory estimation and classification,” 2025 IEEE International Conference on Acoustics, Speech, and Signal Processing, 2025. 7

Pith tools

Reviewed August 11, 2026 · model on record in the stance chip above.