Pith. sign in

REVIEW 4 major objections 6 minor 1 cited by

EFSI-DETR claims that replacing explicit Fourier transforms with a learnable spatial filter bank lets a DETR detector hit 33.1 AP and 24.8 AP-s on VisDrone at 188 FPS, a gain over prior UAV detectors.

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

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

T0 review · deepseek-v4-flash

2026-08-03 07:58 UTC pith:VOJNDTB3

load-bearing objection Workmanlike RT-DETR variant with honest, mostly solid ablations; the SOTA margin over UAV-DETR is uncontrolled and the ablation baseline doesn't match the comparison model, so quote the numbers with a pinch of salt. the 4 major comments →

arxiv 2601.18597 v2 pith:VOJNDTB3 submitted 2026-01-26 cs.CV

EFSI-DETR: Efficient Frequency-Semantic Integration for Real-Time Small Object Detection in UAV Imagery

Yu Xia , Chang Liu , Tianqi Xiang , Zhigang Tu This is my paper
classification cs.CV
keywords real-time object detectionUAV imagerysmall object detectionfrequency-spatial unified synergysemantic feature concentratorDETRfeature pyramidVisDrone
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

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

The paper tries to establish that a DETR-style detector can outperform prior real-time detectors on drone imagery by treating frequency information as learnable spatial filters instead of using explicit Fourier transforms. On VisDrone it reports 33.1 AP and 24.8 small-object AP at 188 FPS with 27.3M parameters, a 1.6 AP and 5.8 AP-s gain over the strongest prior UAV-specific detector, and similar gains on CODrone. A sympathetic reader would care because small objects in aerial scenes are both the most numerous and the hardest to detect, and prior frequency-aware detectors traded accuracy for latency. The paper attributes the gain to three components working together: a dynamic frequency-spatial fusion module, an efficient semantic feature concentrator, and a fine-grained feature retention strategy that changes which pyramid levels reach the decoder.

Core claim

On its own terms, EFSI-DETR claims that small-object detection in UAV imagery can be improved without sacrificing real-time speed. The network simulates a low/mid/high frequency decomposition: average pooling acts as a low-pass branch, an identity path as all-pass, and a depthwise convolution as a learnable high-pass, with content-adaptive weights deciding how much of each band to use. This spectrum is fused through a channel-gating module, while a separate dual-branch module concentrates semantic features using a few expert convolutions and ghost-style cheap features. The third change is to the feature pyramid: shallow high-resolution maps S1 and S2 are fed into the fusion stage, and the co

What carries the argument

The carrying mechanism is DyFusNet, a Dynamic Frequency-Spatial Unified Synergy Network: it decomposes input features into three frequency-band proxies—low via average pooling, mid via identity, high via depthwise convolution—and mixes them with softmax weights computed from global average pooling, then applies channel-wise gating to the aggregated spatial evidence. Around it sit ESFC, which selects among a small set of expert convolutions and uses ghost blocks to keep the semantic branch cheap, and FFR, which redefines the encoder-decoder inputs by keeping S1/S2 and dropping F5. The frequency-band decomposition is what lets the network use spectral cues without paying the latency and memory

Load-bearing premise

The reported margins depend on the comparison being fair: all DETR baselines were retrained with one 'unified hyperparameter configuration' that the paper does not specify, while the strongest prior UAV detector is quoted from its original paper; if that recipe under-trains the baselines, the margin would shrink.

What would settle it

Retrain RT-DETR-R18, DEIM-RT-DETRv2-R18, and UAV-DETR-R18 using exactly the paper's stated training setup (300 epochs, AdamW, learning rate 1e-4, batch size 8, 640x640 input) and evaluate on VisDrone val. If EFSI-DETR's 33.1 AP and 24.8 AP-s do not remain above the best of those baselines by roughly the reported 1.6 AP and 5.8 AP-s margins, the central claim of superiority fails. Repeating with at least three seeds would also show whether the margins exceed run-to-run noise.

Watch this falsifier — get emailed when new claim-graph text bears on it.

If this is right

  • On VisDrone validation, under the paper's protocol, EFSI-DETR reaches 33.1 AP, 52.7 AP50, and 24.8 AP-s at 5.3 ms with 27.3M parameters, exceeding all YOLO and DETR baselines listed.
  • Raising the input size to 800x800 pushes AP to 35.0 and AP-s to 27.3 with the same 5.3 ms latency, so the reported gains persist at higher resolution.
  • The design transfers to CODrone, where EFSI-DETR reaches 20.2 AP versus 17.8 for RT-DETR-R50 and 17.2 for DEIM-RT-DETR-R50.
  • Ablations attribute the largest single jump to the FFR pyramid choice (+4.4 AP), with DyFusNet adding +1.4 AP and ESFC adding +0.4 AP while removing 1.5M parameters.
  • Keeping S1/S2 and dropping F5 improves both accuracy and efficiency, suggesting that for small-object-heavy scenes the coarsest pyramid level contributes more redundancy than useful context.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • The simulated frequency-band approach is not tied to detection; it could be dropped into other latency-critical dense-prediction tasks where full Fourier transforms are too slow but edge and texture cues matter. That extension is not in the paper.
  • The biggest single gain comes from FFR rather than frequency processing, so the visible margin may owe more to feeding high-resolution shallow features and removing the coarse semantic level than to the frequency-spatial module itself.
  • The 5.8-point AP-s gain is concentrated on the hardest scale; a natural test is whether the same recipe helps in large-scene aerial datasets with even more extreme scale distributions, or in oriented object detection where small rotated instances are common.
  • Because all reported numbers come from single runs and DETR training is seed-sensitive, the stability of both the 0.4-point ESFC contribution and the 1.6-point overall margin is unknown; a three-seed replication would show how much of the gain is robust.

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

4 major / 6 minor

Summary. The paper proposes EFSI-DETR, an RT-DETR-based detector for small objects in UAV imagery, with three components: DyFusNet (a frequency-inspired, non-FFT multi-scale fusion module using DMSD and SFCM), ESFC (an efficient semantic feature concentrator using dynamic expert convolution and ghost blocks), and FFR (a fine-grained feature retention strategy that adds shallow features S1/S2 and drops the high-level F5). On VisDrone it reports 33.1 AP / 52.7 AP50 / 24.8 APs at 5.3 ms with 27.3M parameters, and on CODrone 20.2 AP. Ablations show monotone gains from a baseline of 26.9 AP to 33.1 AP.

Significance. If the results hold under a controlled protocol, the paper makes a useful engineering contribution: a hardware-friendly alternative to FFT-based frequency processing and a lightweight semantic concentrator that improve small-object AP on two benchmarks while maintaining real-time throughput. The ablations are monotone and directionally consistent with known small-object sensitivity to high-resolution features. The main weakness is that the evidence as presented does not yet support the SOTA claim because the comparison protocol and the ablation baseline are insufficiently controlled; once these are pinned down, the contribution can be assessed fairly.

major comments (4)
  1. [§IV-B, Table I] The SOTA claim (+1.6 AP, +5.8 AP_s vs UAV-DETR-R50) rests on uncontrolled numbers. UAV-DETR is marked '†' and quoted without retraining, while EFSI and the other DETR baselines are 'retrained under a single unified hyperparameter configuration identical to ours' that is never specified. If the published UAV-DETR numbers use a different recipe (input scale, epochs, augmentation, evaluation), the margin is a comparison artifact. Specify the unified recipe and retrain UAV-DETR under it; report mean±std over at least three seeds.
  2. [§IV-C, Table III / §III-A Eq. (8)] The ablation baseline 'RT-DETR-R18 ε=1' has 25.6M parameters, whereas Table I lists stock RT-DETR-R18 at 19.9M. ε is not defined in the method (Eq. (8) uses e, and ε=1 would send all channels through the frequency path), so the model modified by the ablations is not the same model compared in Table I. Consequently the +6.2 AP total gain and the component attributions are measured from an unspecified, heavier baseline. Add a stock RT-DETR-R18 row with 19.9M params and reconcile the parameter counts.
  3. [§IV-A, Tables III-IV] All headline numbers and ablations are single-run. The ESFC contribution is +0.4 AP and the expert-number variation between K=2 and K=3 is +0.5 AP, both within typical VisDrone seed noise; the DyFusNet gain (+1.4 AP) is also modest. Without seed variance or significance intervals, the incremental claims are not supported. Report mean±std over at least three seeds for Tables I, III, and IV.
  4. [§III-C / §IV-C, Table VI] The FFR design change is not precisely specified. The strategy both adds shallow features S1/S2 and drops the high-level F5, but no description is given of how S1/S2 are projected/fused in the HybridEncoder. Table VI compares 'FFR-F5' with 'FFR' and attributes the 1.2 AP loss to dropping F5, but the parameter count also changes by 2.8M. The two changes are confounded; provide architecture details and split ablations that isolate adding S1/S2 from removing F5.
minor comments (6)
  1. [Fig. 2; §IV-B] Typos: 'Fearute' and 'Internal Fearure' in Fig. 2; 'the the' in §IV-B; 'UA V' spacing throughout.
  2. [Table I] Formatting errors in AP_m/AP_l columns: '43.455.8', '44.044.0', '45.343.3'. Check values.
  3. [Eq. (11)] δ_k are called learned attention weights but their normalization is not specified; state whether they are softmax over experts and whether they are input-dependent.
  4. [Eq. (13)] The adaptive kernel-size formula uses |·|_odd without defining how the odd integer is obtained; also report b and γ values.
  5. [§III-B] The number N of EGBlocks in the residual pathway is never specified; report it in implementation details.
  6. [Abstract / §IV-A] The '188 FPS' is derived from 5.3 ms with TensorRT FP16 on a single RTX 4090; state this context in the abstract or table header to avoid over-generalizing.

Circularity Check

0 steps flagged

No circularity: central claims are benchmark results validated by ablations; protocol concerns are comparison fairness, not definitional reduction.

full rationale

EFSI-DETR's central claim is empirical: AP gains on VisDrone and CODrone measured against public benchmarks and decomposed by ablations (Tables I-III). The equations defining DyFusNet, ESFC, and FFR are trainable network modules; the reported improvements are not obtained by plugging fitted constants back into the same equations (e.g., the learned gates in Eqs. (2) and (6) are optimized weights, not relabeled predictions). No load-bearing self-citation appears: the author-affiliated references [1], [2], and [4] are contextual, and the design choices ('frequency-inspired but non-FFT', 'simulated frequency processing') are explicitly defended by standard signal-processing facts, not by an imported uniqueness theorem. The reviewer-identified concerns about the SOTA comparison are real but belong to experimental protocol, not circularity: UAV-DETR is 'quoted directly from the original papers or official repositories without retraining' (Sec. IV-B), the unified retraining recipe is not spelled out, and the ablation baseline 'RT-DETR-R18 ε=1' is listed at 25.6M parameters (Table III) while Table I lists stock RT-DETR-R18 at 19.9M. These could affect the magnitude of the reported margins, but none is an equation-level reduction of a prediction to its own input, so they do not raise the circularity score.

Axiom & Free-Parameter Ledger

3 free parameters · 5 axioms · 0 invented entities

The headline claims rest on three hand-set hyperparameters (split ratio ϵ=1, K=3 experts, unreported bottleneck ratio r) and five assumptions — the adequacy of the three-operator 'frequency' surrogate, the fairness of the unified baseline-retraining protocol, the FFR prior that shallow features matter more than the coarsest deep map, and validation-set AP as the yardstick — plus the standard Fourier-duality fact. No new entities are postulated: the 'low/mid/high-frequency bands' are disclosed metaphors built from pooling, identity, and depthwise convolution.

free parameters (3)
  • Channel split ratio e (ϵ) in DyFusNet = Reported as ϵ=1 in the ablation protocol (Sec. IV-C); the final-model setting is not explicitly restated
    Eq. 8 splits channels between the frequency path and the bypass; the ratio is a hand-set architectural constant evaluated on the validation set with no sensitivity study.
  • Number of expert convolutions K in DEConv/ESFC = 3
    Table IV sweeps K ∈ {2,3,4,5} on the VisDrone validation set; K=3 is selected because it maximizes AP (33.1) — validation-set model selection.
  • Bottleneck reduction ratio r in SFCM (Eq. 6) = Not stated
    The channel-attention bottleneck W1 ∈ R^{C/r×C} requires a value of r that the paper never reports; it is a hand-set constant affecting the module's capacity.
axioms (5)
  • standard math Spatial convolution corresponds to multiplication in the Fourier domain, so the operator bank {AvgPool3×3, Identity, 3×3 depthwise conv} adequately emulates low/mid/high-band frequency decomposition (Eq. 3).
    Invoked in Sec. III-A 'Why simulated frequency processing'; the Fourier-duality fact is standard, but its use as a surrogate for true spectral decomposition is the paper's modeling premise.
  • domain assumption The three-operator frequency proxy with a content-adaptive gate (Eqs. 1–3) preserves the useful behavior of real spectral processing for small-object detection.
    This is the paper's central design claim; it is justified by analogy and by downstream AP gains, not by measuring the spectral selectivity of the learned operators.
  • domain assumption Retraining every DETR-family baseline under 'a single unified hyperparameter configuration identical to ours' (Sec. IV-B) is at least as favorable to the baselines as their own official recipes.
    Load-bearing for the SOTA margin: if RT-DETR/DEIM are under-tuned by the unified recipe, the +1.6/+5.8 point headline gains overstate the architecture improvement. The recipe itself is never specified.
  • domain assumption For UAV small-object detection, shallow high-resolution features (S1, S2) are the binding constraint and the coarsest deep feature F5 is redundant (FFR, Sec. III-C; Table VI).
    The FFR design and its +4.4 AP ablation gain rest on this empirical prior; note the tension with Table V, where ESFC is most effective at the deep stage, implying deep semantics still matter.
  • domain assumption Validation-set AP on VisDrone/CODrone is an adequate measure of detector performance for the paper's claims.
    Sec. IV-A: 'all models are evaluated on the validation sets.' No test-set results are reported, so the headline numbers inherit any train/val distribution gap.

pith-pipeline@v1.3.0-alltime-deepseek · 14412 in / 35450 out tokens · 336109 ms · 2026-08-03T07:58:29.799464+00:00 · methodology

0 comments
read the original abstract

Real-time small object detection in Unmanned Aerial Vehicle (UAV) imagery remains challenging due to limited feature representation and ineffective multi-scale fusion. Existing methods underutilize frequency information and rely on static convolutional operations, which constrain the capacity to obtain rich feature representations and hinder the effective exploitation of deep semantic features. To address these issues, we propose EFSI-DETR, a novel detection framework that integrates efficient semantic feature enhancement with dynamic frequency-spatial guidance. EFSI-DETR comprises two main components: (1) a Dynamic Frequency-Spatial Unified Synergy Network (DyFusNet) that jointly exploits frequency and spatial cues for robust multi-scale feature fusion, (2) an Efficient Semantic Feature Concentrator (ESFC) that enables deep semantic extraction with minimal computational cost. Furthermore, a Fine-grained Feature Retention (FFR) strategy is adopted to incorporate spatially rich shallow features during fusion to preserve fine-grained details, crucial for small object detection in UAV imagery. Extensive experiments on VisDrone and CODrone benchmarks demonstrate that our EFSI-DETR achieves the state-of-the-art performance with real-time efficiency, yielding improvement of \textbf{1.6}\% and \textbf{5.8}\% in AP and AP$_{s}$ on VisDrone, while obtaining \textbf{188} FPS inference speed on a single RTX 4090 GPU.

Figures

Figures reproduced from arXiv: 2601.18597 by Chang Liu, Tianqi Xiang, Yu Xia, Zhigang Tu.

Figure 1
Figure 1. Figure 1: (a) 3D density distribution of object width and height [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Overview of the proposed EFSI-DETR. DyFusNet exploits frequency and spatial cues to enable robust and adaptive multi [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: Visualization of detection results of RT-DETR and EFSI-DETR. [PITH_FULL_IMAGE:figures/full_fig_p008_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: Visualization of feature maps for variants composed of [PITH_FULL_IMAGE:figures/full_fig_p008_4.png] view at source ↗

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Forward citations

Cited by 1 Pith paper

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

  1. CollabOD: Collaborative Multi-Backbone with Cross-scale Vision for UAV Small Object Detection

    cs.CV 2026-03 conditional novelty 4.0

    CollabOD improves UAV small-object detection via dual-path detail preservation, dense aggregation, bilateral reweighting, and a reparameterized detail-aware head, reporting 52.4 AP50 on VisDrone at 65.5 GFLOPs.

Reference graph

Works this paper leans on

51 extracted references · 5 linked inside Pith · cited by 1 Pith paper

  1. [1]

    Patch similarity self- knowledge distillation for cross-view geo-localization,

    S. Li, M. Hu, X. Xiao, and Z. Tu, “Patch similarity self- knowledge distillation for cross-view geo-localization,” IEEE Trans. on Circuits and Systems for Video Technol- ogy, vol. 34, no. 6, pp. 5091–5103, 2024. JOURNAL OF LATEX CLASS FILES, VOL. 14, NO. 8, AUGUST 2025 9

    2024

  2. [2]

    Multi-scale atten- tion encoder for street-to-aerial image geo-localization,

    S. Li, Z. Tu, Y . Chen, and T. Yu, “Multi-scale atten- tion encoder for street-to-aerial image geo-localization,” CAAI Trans. on Intelligence Technology, vol. 8, no. 1, p. 166–176, 2022

    2022

  3. [3]

    Real-time and accurate uav pedestrian detection for social distancing monitoring in covid-19 pandemic,

    Z. Shao, G. Cheng, J. Ma, Z. Wang, J. Wang, and D. Li, “Real-time and accurate uav pedestrian detection for social distancing monitoring in covid-19 pandemic,” IEEE Trans. on Multimedia, vol. 24, pp. 2069–2083, 2021

    2069

  4. [4]

    Sdpdet: Learning scale-separated dynamic proposals for end-to-end drone- view detection,

    N. Yin, C. Liu, R. Tian, and X. Qian, “Sdpdet: Learning scale-separated dynamic proposals for end-to-end drone- view detection,”IEEE Trans. on Multimedia, vol. 26, p. 7812–7822, 2024

    2024

  5. [5]

    Adazoom: Towards scale- aware large scene object detection,

    J. Xu, Y .-L. Li, and S. Wang, “Adazoom: Towards scale- aware large scene object detection,”IEEE Trans. on Multimedia, vol. 25, p. 4598–4609, 2023

    2023

  6. [6]

    Ex- tended feature pyramid network for small object detec- tion,

    C. Deng, M. Wang, L. Liu, Y . Liu, and J. Yunliang, “Ex- tended feature pyramid network for small object detec- tion,”IEEE Trans. on Multimedia, vol. 24, p. 1968–1979, 2021

    1968

  7. [7]

    Arbitrack: A novel multi-object tracking framework for a moving aav to detect and track arbitrarily oriented targets,

    Y . Chen, J. Wang, Q. Zhou, and H. Hu, “Arbitrack: A novel multi-object tracking framework for a moving aav to detect and track arbitrarily oriented targets,”IEEE Trans. Multimedia, vol. 27, p. 5387–5397, 2025

    2025

  8. [8]

    Bounding box vectorization for oriented object detection with tani- moto coefficient regression,

    L. Wang, Y . Zhan, W. Liu, B. Yu, and D. Tao, “Bounding box vectorization for oriented object detection with tani- moto coefficient regression,”IEEE Trans. on Multimedia, vol. 26, p. 5181–5193, 2024

    2024

  9. [9]

    Feature pyramid networks for object detection,

    T.-Y . Lin, P. Doll ´ar, R. Girshick, K. He, B. Hariharan, and S. Belongie, “Feature pyramid networks for object detection,” inProc. IEEE Conf. Computer Vision and Pattern Recognition, 2017, pp. 936–944

    2017

  10. [10]

    Detrs beat yolos on real-time object detection,

    Y . Zhao, W. Lv, S. Xu, J. Wei, G. Wang, Q. Dang, Y . Liu, and J. Chen, “Detrs beat yolos on real-time object detection,” inProc. IEEE Conf. Computer Vision and Pattern Recognition, 2024, pp. 16 965–16 974

    2024

  11. [11]

    Yolov10: Real-time end-to-end object detection,

    A. Wang, H. Chen, L. Liu, K. Chen, Z. Lin, J. Han, and G. Ding, “Yolov10: Real-time end-to-end object detection,” inProc. Neural Information Proc. Systems, vol. 37, 2024, pp. 107 984–108 011

    2024

  12. [12]

    Ad- yolo: A real-time yolo network with swin transformer and attention mechanism for airport scene detection,

    W. Zhou, C. Cai, C. Li, H. Xu, and H. Shi, “Ad- yolo: A real-time yolo network with swin transformer and attention mechanism for airport scene detection,” IEEE Transactions on Instrumentation and Measurement, vol. 73, pp. 1–12, 2024

    2024

  13. [13]

    Scaled- yolov4: Scaling cross stage partial network,

    C.-Y . Wang, A. Bochkovskiy, and H.-y. Liao, “Scaled- yolov4: Scaling cross stage partial network,” inProc. IEEE Conf. Computer Vision and Pattern Recognition, 2021, pp. 13 024–13 033

    2021

  14. [14]

    Deep convolutional networks on graph-structured data,

    Z. Ge, S. Liu, F. Wang, Z. Li, and J. Sun, “Deep convolutional networks on graph-structured data,”ArXiv, vol. abs/2107.08430, 2021

    Pith/arXiv arXiv 2021

  15. [15]

    Yolov12: Attention-centric real-time object detectors,

    Y . Tian, Q. Ye, and D. Doermann, “Yolov12: Attention-centric real-time object detectors,”ArXiv, vol. abs/2502.12524, 2025

    Pith/arXiv arXiv 2025

  16. [16]

    Attention is all you need,

    A. Vaswani, N. Shazeer, N. Parmar, J. Uszkoreit, L. Jones, A. N. Gomez, L. Kaiser, and I. Polosukhin, “Attention is all you need,” inProc. Neural Information Proc. Systems, 2017, p. 6000–6010

    2017

  17. [17]

    You only look once: Unified, real-time object detection,

    J. Redmon, S. Divvala, R. Girshick, and A. Farhadi, “You only look once: Unified, real-time object detection,” in Proc. IEEE Conf. Computer Vision and Pattern Recog- nition, 2016, pp. 779–788

    2016

  18. [18]

    Fcos: Fully convolutional one-stage object detection,

    Z. Tian, C. Shen, H. Chen, and T. He, “Fcos: Fully convolutional one-stage object detection,” inProc. IEEE Int. Conf. Computer Vision, 2019, pp. 9626–9635

    2019

  19. [19]

    Ultralytics yolov8,

    G. Jocher, A. Chaurasia, and J. Qiu, “Ultralytics yolov8,” 2023. [Online]. Available: https://github.com/ultralytics/ultralytics

    2023

  20. [20]

    Microsoft coco: Common objects in context,

    T.-Y . Lin, M. Maire, S. Belongie, J. Hays, P. Perona, D. Ramanan, P. Doll ´ar, and C. L. Zitnick, “Microsoft coco: Common objects in context,” inProc. European Conf. Computer Vision, 2014, pp. 740–755

    2014

  21. [21]

    Augmentation for small object detection,

    M. Kisantal, Z. Wojna, J. Murawski, J. Naruniec, and K. Cho, “Augmentation for small object detection,” in Proc. International Conf. on Advances in Computing and Information Technology, 2019, pp. 119–133

    2019

  22. [22]

    An analysis of scale invariance in object detection - snip,

    B. Singh and L. S. Davis, “An analysis of scale invariance in object detection - snip,” inProc. IEEE Conf. Computer Vision and Pattern Recognition, 2018, pp. 3578–3587

    2018

  23. [23]

    Clustered object detection in aerial images,

    F. Yang, H. Fan, P. Chu, E. Blasch, and H. Ling, “Clustered object detection in aerial images,” inProc. IEEE Int. Conf. Computer Vision, 2019, pp. 8310–8319

    2019

  24. [24]

    Density map guided object detection in aerial images,

    C. Li, T. Yang, S. Zhu, C. Chen, and S. Guan, “Density map guided object detection in aerial images,” inProc. IEEE Conf. Computer Vision and Pattern Recognition Workshops, 2020, pp. 737–746

    2020

  25. [25]

    Querydet: Cascaded sparse query for accelerating high-resolution small object detection,

    C. Yang, Z. Huang, and N. Wang, “Querydet: Cascaded sparse query for accelerating high-resolution small object detection,” inProc. IEEE Conf. Computer Vision and Pattern Recognition, 2022, pp. 13 658–13 667

    2022

  26. [26]

    Adaptive sparse convolutional networks with global context en- hancement for faster object detection on drone images,

    B. Du, Y . Huang, J. Chen, and D. Huang, “Adaptive sparse convolutional networks with global context en- hancement for faster object detection on drone images,” inProc. IEEE Conf. Computer Vision and Pattern Recog- nition, 2023, pp. 13 435–13 444

    2023

  27. [27]

    Path aggregation network for instance segmentation,

    S. Liu, L. Qi, H. Qin, J. Shi, and J. Jia, “Path aggregation network for instance segmentation,” inProc. IEEE Conf. Computer Vision and Pattern Recognition, 2018, pp. 8759–8768

    2018

  28. [28]

    Efficientdet: Scalable and efficient object detection,

    M. Tan, R. Pang, and Q. V . Le, “Efficientdet: Scalable and efficient object detection,” inProc. IEEE Conf. Computer Vision and Pattern Recognition, 2020, pp. 10 778–10 787

    2020

  29. [29]

    Nas-fpn: Learning scalable feature pyramid architecture for object detec- tion,

    G. Ghiasi, T.-Y . Lin, and Q. V . Le, “Nas-fpn: Learning scalable feature pyramid architecture for object detec- tion,” inProc. IEEE Conf. Computer Vision and Pattern Recognition, 2019, pp. 7029–7038

    2019

  30. [30]

    You only look one-level feature,

    Q. Chen, Y . Wang, T. Yang, X. Zhang, J. Cheng, and J. Sun, “You only look one-level feature,” inProc. IEEE Conf. Computer Vision and Pattern Recognition, 2021, pp. 13 034–13 043

    2021

  31. [31]

    Uav-detr: Efficient end-to-end object detection for unmanned aerial vehicle imagery,

    H. Zhang, K. Liu, Z. Gan, and G.-N. Zhu, “Uav-detr: Efficient end-to-end object detection for unmanned aerial vehicle imagery,”ArXiv, vol. abs/2501.01855, 2025

    Pith/arXiv arXiv 2025

  32. [32]

    A theory for multiresolution signal decom- JOURNAL OF LATEX CLASS FILES, VOL. 14, NO. 8, AUGUST 2025 10 position: the wavelet representation,

    S. Mallat, “A theory for multiresolution signal decom- JOURNAL OF LATEX CLASS FILES, VOL. 14, NO. 8, AUGUST 2025 10 position: the wavelet representation,”IEEE Trans. on Pattern Analysis and Machine Intelligence, vol. 11, no. 7, pp. 674–693, 1989

    2025

  33. [33]

    Drop an octave: Reducing spatial redundancy in convolutional neural net- works with octave convolution,

    Y . Chen, H. Fan, B. Xu, Z. Yan, Y . Kalantidis, M. Rohrbach, S. Yan, and J. Feng, “Drop an octave: Reducing spatial redundancy in convolutional neural net- works with octave convolution,” inProc. IEEE Int. Conf. Computer Vision, 2019, pp. 3434–3443

    2019

  34. [34]

    Making convolutional networks shift- invariant again,

    R. Zhang, “Making convolutional networks shift- invariant again,” inProc. Int. Conf. Machine Learning, 2019, pp. 7324–7334

    2019

  35. [35]

    Spectral represen- tations for convolutional neural networks,

    O. Rippel, J. Snoek, and R. P. Adams, “Spectral represen- tations for convolutional neural networks,” inAdvances in Neural Information Proc. Systems, 2015, pp. 1–10

    2015

  36. [36]

    Fnet: Mixing tokens with fourier transforms,

    J. Lee-Thorp, J. Ainslie, I. Eckstein, and S. Onta ˜n´on, “Fnet: Mixing tokens with fourier transforms,” inProc. Conf. North American Chapter of the Association for Computational Linguistics: Human Language Technolo- gies, 2022, pp. 4296–4313

    2022

  37. [37]

    Harmonic networks: Deep translation and rotation equivariance,

    D. E. Worrall, S. J. Garbin, D. Turmukhambetov, and G. J. Brostow, “Harmonic networks: Deep translation and rotation equivariance,” inProc. IEEE Conf. Computer Vision and Pattern Recognition, 2017, pp. 5028–5037

    2017

  38. [38]

    Ghostnet: More features from cheap operations,

    K. Han, Y . Wang, Q. Tian, J. Guo, C. Xu, and C. Xu, “Ghostnet: More features from cheap operations,” in Proc. IEEE Conf. Computer Vision and Pattern Recog- nition, 2020, pp. 1577–1586

    2020

  39. [39]

    Eca- net: Efficient channel attention for deep convolutional neural networks,

    Q. Wang, B. Wu, P. Zhu, P. Li, W. Zuo, and Q. Hu, “Eca- net: Efficient channel attention for deep convolutional neural networks,” inProc. IEEE Conf. Computer Vision and Pattern Recognition, 2020, pp. 11 531–11 539

    2020

  40. [40]

    Focal loss for dense object detection,

    T.-Y . Lin, P. Goyal, R. Girshick, K. He, and P. Doll ´ar, “Focal loss for dense object detection,” inProc. IEEE Int. Conf. Computer Vision, 2017, pp. 2980–2988

    2017

  41. [41]

    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,”IEEE Trans. Pattern Analysis and Machine Intelligence, vol. 39, no. 6, pp. 1137–1149, 2016

    2016

  42. [42]

    Centernet: Keypoint triplets for object detection,

    K. Duan, S. Bai, L. Xie, H. Qi, Q. Huang, and Q. Tian, “Centernet: Keypoint triplets for object detection,” in Proc. IEEE Int. Conf. Computer Vision, 2019, pp. 6568– 6577

    2019

  43. [43]

    Hrdnet: High- resolution detection network for small objects,

    Z. Liu, G. Gao, L. Sun, and Z. Fang, “Hrdnet: High- resolution detection network for small objects,” inProc. IEEE Int. Conf. Multimedia and Expo, 2021, pp. 1–6

    2021

  44. [44]

    Generalized focal loss: Learning qualified and distributed bounding boxes for dense object detec- tion,

    X. Li, W. Wang, L. Wu, S. Chen, X. Hu, J. Li, J. Tang, and J. Yang, “Generalized focal loss: Learning qualified and distributed bounding boxes for dense object detec- tion,” inAdvances in Neural Information Proc. Systems, 2020, pp. 21 002–21 012

    2020

  45. [45]

    Rtmdet: An empirical study of designing real-time object detectors,

    C. Lyu, W. Zhang, H. Huang, Y . Zhou, Y . Wang, Y . Liu, S. Zhang, and K. Chen, “Rtmdet: An empirical study of designing real-time object detectors,”ArXiv, vol. abs/2212.07784, 2022

    Pith/arXiv arXiv 2022

  46. [46]

    Remdet: Rethinking efficient model design for uav object detec- tion,

    C. Li, R. Zhao, Z. Wang, H. Xu, and X. Zhu, “Remdet: Rethinking efficient model design for uav object detec- tion,” inProc. AAAI Conf. Artificial Intelligence (AAAI), vol. 39, 2025, pp. 4643–4651

    2025

  47. [47]

    Ultralytics yolo11,

    G. Jocher and J. Qiu, “Ultralytics yolo11,” 2024. [On- line]. Available: https://github.com/ultralytics/ultralytics

    2024

  48. [48]

    Deim: Detr with improved matching for fast conver- gence,

    S. Huang, Z. Lu, X. Cun, Y . Yu, X. Zhou, and X. Shen, “Deim: Detr with improved matching for fast conver- gence,” inProc. IEEE Conf. Computer Vision and Pattern Recognition, 2025, pp. 15 162–15 171

    2025

  49. [49]

    Detection and tracking meet drones challenge,

    P. Zhu, L. Wen, D. Du, X. Bian, H. Fan, Q. Hu, and H. Ling, “Detection and tracking meet drones challenge,” IEEE Trans. Pattern Analysis and Machine Intelligence, vol. 44, no. 11, pp. 7380–7399, 2021

    2021

  50. [50]

    More clear, more flexible, more precise: A comprehensive oriented object detection benchmark for uav,

    K. Ye, H. Tang, B. Liu, P. Dai, L. Cao, and R. Ji, “More clear, more flexible, more precise: A comprehensive oriented object detection benchmark for uav,”ArXiv, vol. abs/2504.20032, 2025. Yu Xiareceived the B.S. degree from Northeastern University, Shenyang, China, in 2021, and the M.S. degree from Jiangnan University, Wuxi, China, in

    Pith/arXiv arXiv 2025

  51. [2024]

    degree in the State Key Laboratory of Information Engi- neering in Surveying, Mapping and Remote Sensing at Wuhan University

    He is currently pursuing the Ph.D. degree in the State Key Laboratory of Information Engi- neering in Surveying, Mapping and Remote Sensing at Wuhan University. His research interests include computer vision, image processing, object detection, and human action recognition. Chang Liuis currently working toward the PhD degree with the School of Computer Sc...

    2023