Pith. sign in

REVIEW 2 major objections 5 minor 60 references

A lightweight UAV detector keeps small objects visible by symmetric sampling and scale-aware fusion that switches with model size.

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 · grok-4.5

2026-07-11 21:30 UTC pith:KFQPUSJJ

load-bearing objection Solid engineering paper: two practical modules and a usable mul/add fusion heuristic that actually move the lightweight UAV detector frontier, with clean ablations and open code. the 2 major comments →

arxiv 2607.04125 v1 pith:KFQPUSJJ submitted 2026-07-05 cs.CV

FRFDet: Efficient UAV Small Object Detection with Symmetric Sampling and Scalable Fusion

Yunzhong Si , Huiying Xu , Xinzhong Zhu , Yang Liu , Yao Dong , Wenhao Zhang , Hongbo Li This is my paper
classification cs.CV
keywords small object detectionUAV imageryfeature samplingmulti-scale fusionlightweight detectorVisDronereal-time detection
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.

UAV images are packed with tiny objects that standard detectors lose to background clutter and downsampling. FRFDet claims that two plug-and-play modules fix this without heavy computation. Inverse Bidirectional Sampling rebuilds features with matching expansion-compression and spatial reordering so details survive both down- and up-sampling. Scale-Feature Relationship Cross-Fusion then fuses multi-scale maps differently by model width: multiplication for compact nets, addition for larger ones. On VisDrone, UAVDT, HazyDet and COCO the resulting family of detectors sets new marks among lightweight models while staying small and fast enough for onboard UAV use. A sympathetic reader cares because the same two modules scale cleanly from tiny to large variants and transfer beyond aerial scenes.

Core claim

The paper establishes that small-object detection in UAV imagery improves when feature sampling is made approximately symmetric (via IBS) and when cross-scale fusion is chosen according to model capacity (multiplication for narrow models, addition for wider ones via SFRCF). These two design choices together yield state-of-the-art accuracy among lightweight single-stage detectors at low parameter, FLOP and latency cost.

What carries the argument

Inverse Bidirectional Sampling (IBS) — a learnable expansion-compression unit plus spatial reordering that is inverted for upsampling — and Scale-Feature Relationship Cross-Fusion (SFRCF), which selects inter-group multiplication or addition according to model width.

Load-bearing premise

The claim that multiplication suits compact models and addition suits wider ones is a general capacity effect, not an artifact of the YOLO-based architecture, training recipe, or the three UAV datasets used to observe the trend.

What would settle it

Train a non-YOLO backbone family of matching widths on the same UAV splits and check whether the multiplication-versus-addition preference reverses or disappears; if it does, the scale-dependent fusion rule is architecture-specific rather than general.

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

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

2 major / 5 minor

Summary. The paper proposes FRFDet, a lightweight single-stage detector for UAV small-object detection built on a YOLO11-style pipeline. It introduces two modules: Inverse Bidirectional Sampling (IBS), a learnable symmetric expansion–compression and reorg operator for down- and up-sampling that aims to preserve fine spatial detail and reduce background redundancy (Eqs. 1–2 and Appendix B), and Scale-Feature Relationship Cross-Fusion (SFRCF), a group-wise fusion head that uses inter-group multiplication for compact models and addition for wider ones (Eqs. 3–5, Fig. 5). Experiments on VisDrone, UAVDT, HazyDet, and MS COCO report competitive or SOTA AP among lightweight real-time detectors at modest parameter/FLOP/FPS cost, with component ablations (Table V), residual and kernel controls (Tables VII–VIII, XI), and entropy analyses (Figs. 6–7) supporting the claimed gains.

Significance. If the reported gains hold under independent reimplementation, the work is a solid, practical contribution to resource-constrained UAV detection: two plug-and-play modules, multi-benchmark evidence including adverse weather (HazyDet), transfer to COCO, and open code. The scale-dependent fusion observation (mul for compact, add for wide) is an interesting empirical design heuristic within the YOLO11-derived family and is backed by controlled width/depth sweeps and entropy diagnostics. Strengths include matched-protocol lightweight comparisons, residual/kernel ablations, and public code. The contribution is incremental rather than foundational, but well suited to the applied CV / remote-sensing detection literature.

major comments (2)
  1. Section III.C and Fig. 5 (with Appendix A / Table VI): the claim that multiplicative fusion is preferred for compact models and additive for larger ones is presented as a capacity-dependent phenomenon driven by quadratic interaction growth (Eqs. 4–5). All scaling evidence is internal to the FRFDet/YOLO11-derived family (fixed depth 0.5, width 0.25–1.5). The paper should either (i) state this scope limitation explicitly in the abstract and conclusion, or (ii) add at least one external backbone/neck control (e.g., a non-YOLO lightweight detector) so the design rule is not over-generalized beyond the architecture family used to discover it. This does not overturn the SOTA tables but is load-bearing for the interpretive claim about SFRCF.
  2. Tables I–IV and implementation details (§IV.B): several SOTA comparisons mix 640×640 YOLO-style models with high-resolution two-stage or multi-scale methods (1333×800, 2666×1600) and cluster-cropped settings. While the paper marks “o / ca / dc” and reports its own high-res and cropped variants, a short protocol-matched subset (same input size, same training epochs, same backbone family) would make the “SOTA among lightweight detectors” claim cleaner and reduce the risk that gains are partly protocol-driven. The core lightweight numbers (FRFDet-T/S vs YOLO11/RemDet) already look solid; the request is for clearer isolation rather than new experiments that change the conclusion.
minor comments (5)
  1. Fig. 2 and §III.B: the “inverse” / bidirectional symmetry of IBS is clear from the red arrows and Eqs. 1–2, but the residual asymmetry (residual kept in IBS-D, removed in IBS-U; Table VIII) should be stated once in the main text, not only in the appendix.
  2. Table I: FPS for several competitors is missing (“–”); either measure under the same RTX 4090 / batch-1 protocol or drop FPS from those rows to avoid incomplete efficiency claims.
  3. Notation: C_in / C_mid / C_out and the expansion ratio r appear in Appendix B but are only lightly introduced in the main text; a one-line definition near Eq. (1) would help.
  4. Typos / consistency: “UA V” spacing, “overlooks” (III.C), and occasional “from-scratch” vs “from scratch” should be cleaned for camera-ready.
  5. Appendix D entropy analysis is valuable; a brief pointer in §IV.D that the entropy gap supports the redundancy-suppression claim would better connect the appendix to the main narrative.

Circularity Check

0 steps flagged

No significant circularity: empirical architecture proposal with independent multi-benchmark measurements

full rationale

FRFDet is a standard empirical CV paper proposing two modules (IBS via symmetric expansion-compression + reorganization, SFRCF via scale-dependent mul/add fusion) and evaluating them on public external benchmarks (VisDrone val, UAVDT, HazyDet, MS COCO). Performance claims (SOTA among lightweight detectors, ablations in Table V, scaling in Fig. 5/Appendix A, entropy analysis in Figs. 6-7) are measured quantities, not algebraic consequences of fitted inputs or self-defined quantities. The mul-for-compact / add-for-wide rule is an observation from controlled width/depth sweeps inside the YOLO11-derived family, then adopted as a design choice; it is not presented as a first-principles derivation that forces the reported APs. Self-citations (e.g., RemDet) appear only as comparison baselines, not as load-bearing uniqueness theorems or smuggled ansatze. No equation reduces to its own input by construction, no parameter fitted on a subset is relabeled a prediction of a related quantity, and no central claim rests solely on an unverified self-citation chain. The derivation chain is therefore self-contained against external data.

Axiom & Free-Parameter Ledger

3 free parameters · 3 axioms · 2 invented entities

As an empirical architecture paper the central claims rest on a handful of design choices (expansion ratio, residual placement, mul/add switch) that are selected by ablation rather than derived, plus standard domain facts about UAV imagery. No new physical entities are postulated; the invented modules are engineering constructs whose value is measured by external benchmarks.

free parameters (3)
  • IBS expansion ratio r = 2
    Set to 2 after ablation (Table VII) for best accuracy-FLOPs trade-off; other values (1,4) are inferior.
  • width and depth scaling factors = see Table VI
    Hand-chosen per model size (Table VI: T 0.25/0.5, S 0.5/0.5, M 1.0/0.5, L 1.0/1.0, X 1.5/0.5) to produce the reported parameter/FLOP counts.
  • fusion operator choice (mul vs add) = mul for T/S, add for L/X
    Selected per model capacity after observing Fig. 5; not a continuous fitted weight but a discrete hyper-parameter that the central claim depends on.
axioms (3)
  • domain assumption Objects smaller than 32×32 pixels are progressively degraded by conventional down-sampling and occupy <8 % of UAV image area.
    Stated in Introduction and used to motivate both IBS and the redundancy-suppression goal; taken from VisDrone statistics [16].
  • ad hoc to paper Inter-group multiplicative fusion produces an interaction space that grows as C_in² while additive fusion grows linearly.
    Section III.C; used to justify why mul helps compact models but hurts wider ones. The quadratic claim is plausible but not formally derived from first principles inside the paper.
  • standard math Shannon entropy of discretized feature values measures foreground saliency versus background redundancy.
    Appendix D; standard information-theoretic tool applied post-hoc to interpret IBS/SFRCF.
invented entities (2)
  • Inverse Bidirectional Sampling (IBS) no independent evidence
    purpose: Symmetric learnable down-/up-sampling that expands-compresses channels and reorders space to preserve fine details while suppressing background.
    Core architectural novelty; independent_evidence=false because its value is demonstrated only by the paper’s own ablations and entropy plots.
  • Scale-Feature Relationship Cross-Fusion (SFRCF) no independent evidence
    purpose: Group-wise cross-scale fusion that switches between element-wise multiplication and addition according to model width.
    Second core novelty; again validated only inside the FRFDet experiments.

pith-pipeline@v1.1.0-grok45 · 25236 in / 2834 out tokens · 32089 ms · 2026-07-11T21:30:21.727967+00:00 · methodology

0 comments
read the original abstract

Small object detection in Unmanned Aerial Vehicle (UAV) imagery remains challenging under adverse conditions, including complex weather, low illumination, and sensor noise. These challenges mainly stem from severe background clutter, fine-grained detail degradation, and suboptimal semantic-spatial feature fusion, which jointly hinder robust small-object representation. To this end, we propose FRFDet, a lightweight yet effective single-stage detector tailored for UAV-based small object detection. FRFDet proposes two plug-and-play modules: Inverse Bidirectional Sampling (IBS) and Scale-Feature Relationship Cross-Fusion (SFRCF). IBS preserves critical spatial details via channel expansion-compression and bidirectional pattern reconstruction, improving feature alignment. SFRCF explicitly models scale-dependent fusion behaviors, revealing that inter-group element-wise multiplication favors compact models, while inter-group additive fusion benefits larger architectures. Extensive experiments on VisDrone, UAVDT, HazyDet, and MS COCO demonstrate that FRFDet achieves state-of-the-art performance among lightweight detectors with low computational cost, compact parameters, and fast inference, making it well suited for resource-constrained UAV platforms.

Figures

Figures reproduced from arXiv: 2607.04125 by Hongbo Li, Huiying Xu, Wenhao Zhang, Xinzhong Zhu, Yang Liu, Yao Dong, Yunzhong Si.

Figure 1
Figure 1. Figure 1: Feature map comparison. Standard sampling shows [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Overview of FRFDet. IBS enables symmetric down-/up-sampling, and SFRCF performs scale-aware cross-fusion to [PITH_FULL_IMAGE:figures/full_fig_p002_2.png] view at source ↗
Figure 4
Figure 4. Figure 4: Fusion visualizations on VisDrone. Naive fusion shows [PITH_FULL_IMAGE:figures/full_fig_p003_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: Comparison of AP at different scales of FRFDet on [PITH_FULL_IMAGE:figures/full_fig_p003_5.png] view at source ↗
Figure 6
Figure 6. Figure 6: Quantitative visualization of foreground and background entropy for FRFDet’s IBS compared to the baseline model [PITH_FULL_IMAGE:figures/full_fig_p010_6.png] view at source ↗
Figure 7
Figure 7. Figure 7: Quantitative visualization of foreground and background entropy for the SFRCF module in our FRFDet compared to [PITH_FULL_IMAGE:figures/full_fig_p011_7.png] view at source ↗
Figure 8
Figure 8. Figure 8: Visualization of detection results on VisDrone, com [PITH_FULL_IMAGE:figures/full_fig_p011_8.png] view at source ↗
Figure 9
Figure 9. Figure 9: Visualization of detection results on HazyDet’s RDDTS, comparing our FRFDet-T with state-of-the-art real-time [PITH_FULL_IMAGE:figures/full_fig_p012_9.png] view at source ↗

discussion (0)

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

Reference graph

Works this paper leans on

60 extracted references · 9 linked inside Pith

  1. [1]

    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 transactions on pattern analysis and machine intelligence, vol. 39, no. 6, pp. 1137–1149, 2016

    2016

  2. [2]

    Yolov7: Trainable bag-of-freebies sets new state-of-the-art for real-time object detectors,

    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 detectors,” inProceedings of the IEEE/CVF conference on computer vision and pattern recognition, 2023, pp. 7464–7475

    2023

  3. [3]

    An image is worth 16x16 words: Transformers for image recognition at scale,

    A. Dosovitskiy, “An image is worth 16x16 words: Transformers for image recognition at scale,”arXiv preprint arXiv:2010.11929, 2020

    Pith/arXiv arXiv 2010

  4. [4]

    Swin transformer: Hierarchical vision transformer using shifted windows,

    Z. Liu, Y . Lin, Y . Cao, H. Hu, Y . Wei, Z. Zhang, S. Lin, and B. Guo, “Swin transformer: Hierarchical vision transformer using shifted windows,” inProceedings of the IEEE/CVF international conference on computer vision, 2021, pp. 10 012–10 022

    2021

  5. [5]

    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,” inProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2024, pp. 16 965–16 974

    2024

  6. [6]

    Mamba: Linear-time sequence modeling with selective state spaces,

    A. Gu and T. Dao, “Mamba: Linear-time sequence modeling with selective state spaces,”arXiv preprint arXiv:2312.00752, 2023

    Pith/arXiv arXiv 2023

  7. [7]

    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,” inComputer Vision–ECCV 2014: 13th European Conference, Zurich, Switzerland, September 6-12, 2014, Proceedings, Part V 13. Springer, 2014, pp. 740–755

    2014

  8. [8]

    Refining activation down- sampling with softpool,

    A. Stergiou, R. Poppe, and G. Kalliatakis, “Refining activation down- sampling with softpool,” inProceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), October 2021, pp. 10 357– 10 366

    2021

  9. [9]

    Feature selective anchor-free module for single-shot object detection,

    C. Zhu, Y . He, and M. Savvides, “Feature selective anchor-free module for single-shot object detection,” inProceedings of the IEEE/CVF conference on computer vision and pattern recognition, 2019, pp. 840– 849

    2019

  10. [10]

    Adaptive sparse convolutional networks with global context enhancement for faster object detection on drone images,

    B. Du, Y . Huang, J. Chen, and D. Huang, “Adaptive sparse convolutional networks with global context enhancement for faster object detection on drone images,” inProceedings of the IEEE/CVF conference on computer vision and pattern recognition, 2023, pp. 13 435–13 444

    2023

  11. [11]

    Remdet: Rethinking efficient model design for uav object detection,

    C. Li, R. Zhao, Z. Wang, H. Xu, and X. Zhu, “Remdet: Rethinking efficient model design for uav object detection,” inProceedings of the AAAI Conference on Artificial Intelligence, vol. 39, no. 5, 2025, pp. 4643–4651

    2025

  12. [12]

    Slicing aided hyper inference and fine-tuning for small object detection,

    F. C. Akyon, S. O. Altinuc, and A. Temizel, “Slicing aided hyper inference and fine-tuning for small object detection,” in2022 IEEE international conference on image processing (ICIP). IEEE, 2022, pp. 966–970

    2022

  13. [13]

    Ufpmp-det: Toward accurate and efficient object detection on drone imagery,

    Y . Huang, J. Chen, and D. Huang, “Ufpmp-det: Toward accurate and efficient object detection on drone imagery,” inProceedings of the AAAI conference on artificial intelligence, vol. 36, no. 1, 2022, pp. 1026–1033

    2022

  14. [14]

    A global-local self-adaptive network for drone-view object detection,

    S. Deng, S. Li, K. Xie, W. Song, X. Liao, A. Hao, and H. Qin, “A global-local self-adaptive network for drone-view object detection,” IEEE Transactions on Image Processing, vol. 30, pp. 1556–1569, 2020

    2020

  15. [15]

    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,” in2021 IEEE international conference on multimedia and expo (ICME). IEEE, 2021, pp. 1–6

    2021

  16. [16]

    Visdrone-det2019: The vision meets drone object detection in image challenge results,

    D. Du, P. Zhu, L. Wen, X. Bian, H. Lin, Q. Hu, T. Peng, J. Zheng, X. Wang, Y . Zhanget al., “Visdrone-det2019: The vision meets drone object detection in image challenge results,” inProceedings of the IEEE/CVF international conference on computer vision workshops, 2019, pp. 0–0

    2019

  17. [17]

    The unmanned aerial vehicle benchmark: Object detection and tracking,

    D. Du, Y . Qi, H. Yu, Y . Yang, K. Duan, G. Li, W. Zhang, Q. Huang, and Q. Tian, “The unmanned aerial vehicle benchmark: Object detection and tracking,” inProceedings of the European conference on computer vision (ECCV), 2018, pp. 370–386

    2018

  18. [18]

    Hazydet: Open-source benchmark for drone- view object detection with depth-cues in hazy scenes,

    C. Feng, Z. Chen, R. Kou, G. Gao, C. Wang, X. Li, X. Shu, Y . Dai, Q. Fu, and J. Yang, “Hazydet: Open-source benchmark for drone- view object detection with depth-cues in hazy scenes,”arXiv preprint arXiv:2409.19833, 2024

    Pith/arXiv arXiv 2024

  19. [19]

    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,” inProceedings of the IEEE international conference on computer vision, 2017, pp. 2980–2988

    2017

  20. [20]

    Bridging the gap between anchor-based and anchor-free detection via adaptive training sample selection,

    S. Zhang, C. Chi, Y . Yao, Z. Lei, and S. Z. Li, “Bridging the gap between anchor-based and anchor-free detection via adaptive training sample selection,” inProceedings of the IEEE/CVF conference on computer vision and pattern recognition, 2020, pp. 9759–9768

    2020

  21. [21]

    Disentangle your dense object detector,

    Z. Chen, C. Yang, Q. Li, F. Zhao, Z.-J. Zha, and F. Wu, “Disentangle your dense object detector,” inProceedings of the 29th ACM interna- tional conference on multimedia, 2021, pp. 4939–4948

    2021

  22. [22]

    Reppoints: Point set representation for object detection,

    Z. Yang, S. Liu, H. Hu, L. Wang, and S. Lin, “Reppoints: Point set representation for object detection,” inProceedings of the IEEE/CVF international conference on computer vision, 2019, pp. 9657–9666

    2019

  23. [23]

    Fcos: Fully convolutional one- stage object detection,

    Z. Tian, C. Shen, H. Chen, and T. He, “Fcos: Fully convolutional one- stage object detection,” inProceedings of the IEEE/CVF international conference on computer vision, 2019, pp. 9627–9636

    2019

  24. [24]

    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,” inProceedings of the IEEE/CVF international conference on computer vision, 2019, pp. 6569–6578

    2019

  25. [25]

    Generalized focal loss: Learning qualified and distributed bounding boxes for dense object detection,

    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 detection,”Advances in neural information processing systems, vol. 33, pp. 21 002–21 012, 2020

    2020

  26. [26]

    Tood: Task- aligned one-stage object detection,

    C. Feng, Y . Zhong, Y . Gao, M. R. Scott, and W. Huang, “Tood: Task- aligned one-stage object detection,” in2021 IEEE/CVF International Conference on Computer Vision (ICCV). IEEE Computer Society, 2021, pp. 3490–3499

    2021

  27. [27]

    End-to-end object detection with transformers,

    N. Carion, F. Massa, G. Synnaeve, N. Usunier, A. Kirillov, and S. Zagoruyko, “End-to-end object detection with transformers,” in European conference on computer vision. Springer, 2020, pp. 213– 229

    2020

  28. [28]

    Deformable detr: Deformable transformers for end-to-end object detection,

    X. Zhu, W. Su, L. Lu, B. Li, X. Wang, and J. Dai, “Deformable detr: Deformable transformers for end-to-end object detection,”arXiv preprint arXiv:2010.04159, 2020

    Pith/arXiv arXiv 2010

  29. [29]

    Yolov6 v3. 0: A full-scale reloading,

    C. Li, L. Li, Y . Geng, H. Jiang, M. Cheng, B. Zhang, Z. Ke, X. Xu, and X. Chu, “Yolov6 v3. 0: A full-scale reloading,”arXiv preprint arXiv:2301.05586, 2023

    Pith/arXiv arXiv 2023

  30. [30]

    Yolov9: Learning what you want to learn using programmable gradient information,

    C.-Y . Wang, I.-H. Yeh, and H.-Y . Mark Liao, “Yolov9: Learning what you want to learn using programmable gradient information,” inEuro- pean conference on computer vision. Springer, 2024, pp. 1–21

    2024

  31. [31]

    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,” inAdvances in Neural Information Processing Systems, A. Globerson, L. Mackey, D. Belgrave, A. Fan, U. Paquet, J. Tomczak, and C. Zhang, Eds., vol. 37. Curran Associates, Inc., 2024, pp. 107 984–108 011. [Online]. Available: https://proceedings...

    2024

  32. [32]

    Ultralytics yolo11,

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

    2024

  33. [33]

    Yolov12: Attention-centric real-time object detectors,

    Y . Tian, Q. Ye, and D. Doermann, “Yolov12: Attention-centric real-time object detectors,”arXiv preprint arXiv:2502.12524, 2025

    Pith/arXiv arXiv 2025

  34. [34]

    Gold- yolo: Efficient object detector via gather-and-distribute mechanism,

    C. Wang, W. He, Y . Nie, J. Guo, C. Liu, Y . Wang, and K. Han, “Gold- yolo: Efficient object detector via gather-and-distribute mechanism,” Advances in Neural Information Processing Systems, vol. 36, 2024

    2024

  35. [35]

    Hyper-yolo: When visual object detection meets hypergraph computation,

    Y . Feng, J. Huang, S. Du, S. Ying, J.-H. Yong, Y . Li, G. Ding, R. Ji, and Y . Gao, “Hyper-yolo: When visual object detection meets hypergraph computation,”IEEE Transactions on Pattern Analysis and Machine Intelligence, 2024

    2024

  36. [36]

    Clustered object de- tection in aerial images,

    F. Yang, H. Fan, P. Chu, E. Blasch, and H. Ling, “Clustered object de- tection in aerial images,” inProceedings of the IEEE/CVF international conference on computer vision, 2019, pp. 8311–8320

    2019

  37. [37]

    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,” inproceedings of the IEEE/CVF conference on computer vision and pattern recognition workshops, 2020, pp. 190–191

    2020

  38. [38]

    Coarse-grained density map guided object detection in aerial images,

    C. Duan, Z. Wei, C. Zhang, S. Qu, and H. Wang, “Coarse-grained density map guided object detection in aerial images,” inProceedings of the IEEE/CVF International Conference on Computer Vision, 2021, pp. 2789–2798

    2021

  39. [39]

    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,” inProceedings of the IEEE/CVF Conference on computer vision and pattern recognition, 2022, pp. 13 668–13 677

    2022

  40. [40]

    Esod: Efficient small object detection on high-resolution images,

    K. Liu, Z. Fu, S. Jin, Z. Chen, F. Zhou, R. Jiang, Y . Chen, and J. Ye, “Esod: Efficient small object detection on high-resolution images,”IEEE Transactions on Image Processing, 2024

    2024

  41. [41]

    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,” inProceedings of the IEEE conference on computer vision and pattern recognition, 2017, pp. 2117–2125

    2017

  42. [42]

    Efficientdet: Scalable and efficient object detection,

    M. Tan, R. Pang, and Q. V . Le, “Efficientdet: Scalable and efficient object detection,” inProceedings of the IEEE/CVF conference on computer vision and pattern recognition, 2020, pp. 10 781–10 790

    2020

  43. [43]

    Ultralytics yolov8,

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

    2023

  44. [44]

    Dtssnet: Dynamic training sample selection network for uav object detection,

    L. Chen, C. Liu, W. Li, Q. Xu, and H. Deng, “Dtssnet: Dynamic training sample selection network for uav object detection,”IEEE Transactions on Geoscience and Remote Sensing, 2024

    2024

  45. [45]

    Yolc: You only look clusters for tiny object detection in aerial images,

    C. Liu, G. Gao, Z. Huang, Z. Hu, Q. Liu, and Y . Wang, “Yolc: You only look clusters for tiny object detection in aerial images,”IEEE Transactions on Intelligent Transportation Systems, 2024

    2024

  46. [46]

    A lightweight fusion strategy with enhanced inter-layer feature correlation for small object detection,

    Y . Xiao, T. Xu, X. Yu, Y . Fang, and J. Li, “A lightweight fusion strategy with enhanced inter-layer feature correlation for small object detection,” IEEE Transactions on Geoscience and Remote Sensing, 2024

    2024

  47. [47]

    Global-local fusion with semantic information-guidance for accurate small object detection in uav aerial images,

    Y . Chen, Z. Ye, H. Sun, T. Gong, S. Xiong, and X. Lu, “Global-local fusion with semantic information-guidance for accurate small object detection in uav aerial images,”IEEE Transactions on Geoscience and Remote Sensing, 2025

    2025

  48. [48]

    Explaining neural scaling laws,

    Y . Bahri, E. Dyer, J. Kaplan, J. Lee, and U. Sharma, “Explaining neural scaling laws,”Proceedings of the National Academy of Sciences, vol. 121, no. 27, p. e2311878121, 2024

    2024

  49. [49]

    Conditional-pooling for improved data transmission,

    E. Bayraktar and C. B. Yigit, “Conditional-pooling for improved data transmission,”Pattern Recognition, p. 109978, 2023. [Online]. Available: https://www.sciencedirect.com/science/article/pii/ S0031320323006763

    2023

  50. [50]

    Carafe: Content-aware reassembly of features,

    J. Wang, K. Chen, R. Xu, Z. Liu, C. C. Loy, and D. Lin, “Carafe: Content-aware reassembly of features,” inProceedings of the IEEE/CVF international conference on computer vision, 2019, pp. 3007–3016

    2019

  51. [51]

    Learning to upsample by learning to sample,

    W. Liu, H. Lu, H. Fu, and Z. Cao, “Learning to upsample by learning to sample,” inProceedings of the IEEE/CVF International Conference on Computer Vision, 2023, pp. 6027–6037

    2023

  52. [52]

    Afpn: Asymptotic feature pyramid network for object detection,

    G. Yang, J. Lei, Z. Zhu, S. Cheng, Z. Feng, and R. Liang, “Afpn: Asymptotic feature pyramid network for object detection,” in2023 IEEE International Conference on Systems, Man, and Cybernetics (SMC). IEEE, 2023, pp. 2184–2189

    2023

  53. [53]

    Centralized feature pyramid for object detection,

    Y . Quan, D. Zhang, L. Zhang, and J. Tang, “Centralized feature pyramid for object detection,”IEEE Transactions on Image Processing, 2023

    2023

  54. [54]

    Qwen3-vl technical report,

    S. Bai, Y . Cai, R. Chen, K. Chen, X. Chen, Z. Cheng, L. Deng, W. Ding, C. Gao, C. Ge, W. Ge, Z. Guo, Q. Huang, J. Huang, F. Huang, B. Hui, S. Jiang, Z. Li, M. Li, M. Li, K. Li, Z. Lin, J. Lin, X. Liu, J. Liu, C. Liu, Y . Liu, D. Liu, S. Liu, D. Lu, R. Luo, C. Lv, R. Men, L. Meng, X. Ren, X. Ren, S. Song, Y . Sun, J. Tang, J. Tu, J. Wan, P. Wang, P. Wang,...

    Pith/arXiv arXiv 2025

  55. [55]

    Glm-4.1 v-thinking: Towards versatile mul- timodal reasoning with scalable reinforcement learning,

    W. Hong, W. Yu, X. Gu, G. Wang, G. Gan, H. Tang, J. Cheng, J. Qi, J. Ji, L. Panet al., “Glm-4.1 v-thinking: Towards versatile mul- timodal reasoning with scalable reinforcement learning,”arXiv preprint arXiv:2507.01006, 2025

    Pith/arXiv arXiv 2025

  56. [56]

    Perceive anything: Recognize, explain, cap- tion, and segment anything in images and videos,

    W. Lin, X. Wei, R. An, T. Ren, T. Chen, R. Zhang, Z. Guo, W. Zhang, L. Zhang, and H. Li, “Perceive anything: Recognize, explain, cap- tion, and segment anything in images and videos,”arXiv preprint arXiv:2506.05302, 2025

    Pith/arXiv arXiv 2025

  57. [57]

    Personalq: Select, quantize, and serve personalized diffusion models for efficient inference,

    Q. Wang, Q. Guo, Y . Sun, J. Yang, D. Zhang, S. Pang, and Q. Guo, “Personalq: Select, quantize, and serve personalized diffusion models for efficient inference,”arXiv preprint arXiv:2603.22943, 2026. A. ARCHITECTURALSCALINGANALYSIS: WIDTH AND DEPTHFACTORS We construct FRFDet models at different scales by adjusting the width and depth factors, as summarize...

    arXiv 2026

  58. [58]

    Discretization into histogram bins: Divide the value range ofxintoKequally spaced binsB 1, B2, ..., BK, and compute the bin countsc: cj = #{x i ∈B j}, j= 1,2, ..., K(18)

  59. [59]

    Probability normalization: pi = ci PK j=1 cj +ϵ , ϵ >0(19) Here, we setϵto1×10 −9

  60. [60]

    LayerGroup (IBS- D+IBS-U)

    Entropy computation: H(x) =− KX i=1 pi ·log 2(pi), pi >0(20) 1 3 5 7 0.5 0.8 1.1 1.4 1.7 2.0 2.3 2.6 2.9 3.2 3.5 3.8 Foreground Entropy 3.34 2.33 0.60 Entropy 1.32 2.23 1.33 3.21 0.61 11 14 17 20 2.36 0.60 1.33 1.58 1.32 0.61 1.34 0.81 Backbone Neck IBS-D IBS-U IBS-D 1 3 5 7 3.0 3.2 3.4 3.6 3.8 4.0 4.2 4.4 4.6 4.8 5.0 5.2 Background Entropy 3.54 3.89 4.26...