Pith. sign in

REVIEW 5 major objections 5 minor 1 cited by

MGDFIS: Multi-scale Global-detail Feature Integration Strategy for Small Object Detection

T0 review · 5 major / 5 minor · reviewed 2026-08-07 · deepseek-v4-flash

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

Pith's one-line read MGDFIS, a neck-level fusion strategy coupling global context, local detail, and pixel reweighting, lifts small-object detection across one-stage, two-stage, and end-to-end detectors.

desk verdict A plausible method buried under sloppy reporting: the abstract promises experiments the manuscript never shows, but the core fusion strategy is a new combination worth a careful look. read the letter →

arxiv 2506.12697 v4 pith:XYEFKJF5 submitted 2025-06-15 cs.CV cs.AIcs.LG

Yuxiang Wang , Xuecheng Bai , Chuanzhi Xu , Ying Zhou , Weidong Cai This is my paper · ORCID
classification cs.CVcs.AIcs.LG
keywords smallobjectdetectionUAVimagerymulti-scalefeaturefusionattentionmechanismglobal-localintegrationVisDroneDynamicPixelpyramidnetwork
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

Small objects in UAV imagery are easy to lose: they occupy few pixels, have low signal-to-noise ratios, and get blurred by repeated downsampling. This paper proposes MGDFIS, a fuse-at-the-neck strategy that replaces plain concatenation-based cross-scale fusion with three coordinated modules—spectral-spatial attention, global-detail integration, and pixel-level foreground-background reweighting—so that global context and local detail reinforce each other. The paper's claim is that this coupling consistently raises small-object detection accuracy across one-stage, two-stage, and transformer detectors on VisDrone, with the largest reported gain a 12.5-point AP50 jump for Faster R-CNN and a 1.5-point AP50 gain for YOLO11s. If the claim holds, the strategy offers a practical, computationally light way to improve small-object detection on resource-constrained aerial platforms.

What carries the argument

The load-bearing mechanism is the MGDFIS neck, a three-module arrangement inserted at the feature-fusion level. FusionLock-TSS Attention (FTSSA) combines Token Statistics Self-Attention with DynamicTanh scaling, Mona's multi-scale depthwise convolutions, and a Spectral Enhanced Feed-Forward block that reweights features in the Fourier domain; its job is to stabilize spectral-spatial responses at linear complexity. The Global-detail Integration Module (GDIM) runs two blocks in series: a Group Mixing Module (GMM) that slices and recombines the feature map along rows and columns so distant pixels interact within a local receptive field, and a Detail Mixing Module (DMM) that uses 4x6 and 6x4 directional convolutions plus an attention path to recover local shape and texture. Dynamic Pixel Attention Module (DPAM) concatenates the aggregated and refined features, applies a 7x7 convolution and sigmoid, and produces a pixel-wise weight map that rebalances foreground against background before weighted fusion.

What would settle it

Train Faster R-CNN with and without MGDFIS on VisDrone using the released code and the paper's 400-epoch AdamW protocol and check whether AP50 reproduces 53.2% versus the 40.7% baseline; in parallel, inspect the TSSA implementation against Eq. (7), because the printed formula uses symbols $F_{\mathrm{Stos}}$ and $\pi$ that are not defined, so a mismatch would mean the results cannot be attributed to the described module.

Watch

Extended reading notes

Core claim

The paper's central discovery is that the failure mode of standard multi-scale fusion in small-object detection is not only missing context but also lost detail, and both can be addressed in one neck-level design. Concretely, it claims that MGDFIS—built from FusionLock-TSS Attention (FTSSA), a Global-detail Integration Module (GDIM), and a Dynamic Pixel Attention Module (DPAM)—outperforms the corresponding baselines when inserted into SSD, YOLO variants, Faster R-CNN, DETR, and RT-DETR. On VisDrone, Faster R-CNN + MGDFIS reaches 33.4% mAP and 53.2% AP50 versus 31.4% and 40.7% for the baseline, while YOLO11s + MGDFIS reaches 23.5% mAP and 39.4% AP50 versus 22.4% and 37.9%. The abstract additionally reports a YOLO26m + MGDFIS configuration improving AP50:95 from 25.7 to 30.2 and AP50 from 37.2 to 44.2 at 96.1 GFLOPs, and the paper reports 38.9 AP50 and 21.9 AP50:95 on UAVDT and 97.4 AP50 on CARPK. The gains are attributed to synergistic global long-range mixing, local detail refinement, and pixel-wise recalibration rather than to any single module.

Load-bearing premise

The gains assume the formulas in Section III describe the released code exactly, and that every baseline was trained and evaluated under the same protocol, so the 12.5-point AP50 jump for Faster R-CNN is a real effect rather than an artifact of the comparison.

Editorial extensions

If this is right

  • Inserting MGDFIS into Faster R-CNN raises VisDrone AP50 from 40.7% to 53.2%, a 12.5-point gain, while raising mAP from 31.4% to 33.4%.
  • YOLO11s gains 1.1 points in mAP (22.4% to 23.5%) and 1.5 points in AP50 (37.9% to 39.4%), and similar gains appear across YOLOv5s through YOLOv10s.
  • The strategy transfers to transformer-based detectors: DETR and RT-DETR improve in both mAP and AP50, with RT-DETR reaching 29.1% mAP and 47.9% AP50.
  • On datasets outside VisDrone, MGDFIS reports 38.9 AP50 and 21.9 AP50:95 on UAVDT and 97.4 AP50 on CARPK, indicating it is not tuned to a single benchmark.
  • Computational cost stays moderate: in the YOLO11s ablation, full MGDFIS adds 28.2 GFLOPs over the 21.7 GFLOPs baseline, which the paper argues is compatible with real-time UAV use.

Reading between the lines

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

  • Editorial inference: The uneven gains across detectors—12.5 AP50 points for Faster R-CNN versus 0.6–1.5 points for YOLO variants—suggest MGDFIS interacts strongly with two-stage region-proposal pipelines; testing it on Cascade R-CNN or Mask R-CNN would probe whether the effect is structural or specific to Faster R-CNN.
  • Editorial inference: The abstract headline results (YOLO26m: 25.7 to 30.2 AP50:95, 37.2 to 44.2 AP50) do not appear in the full-text tables, which center on YOLO11s; reproducing the YOLO26m configuration would reconcile the two versions of the claim.
  • Editorial inference: Because FTSSA is assembled from previously published components (TSSA, DynamicTanh, SEFF, Mona), an ablation that swaps each component for its standard counterpart would clarify whether the gain comes from the specific combination or from the added parameters and compute.
  • Editorial inference: A natural testable extension is to apply MGDFIS to non-aerial small-object benchmarks such as COCO-mini or SODA to see whether the global-detail coupling transfers beyond UAV imagery.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

5 major / 5 minor

Summary. The paper proposes MGDFIS, a neck-level multi-scale feature-fusion strategy for small-object detection in UAV imagery. It combines three modules: FusionLock-TSS Attention (FTSSA), a Global-Detail Integration Module (GDIM) built from a Global Mixing Module (GMM) and a Detail Mixing Module (DMM), and a Dynamic Pixel Attention Module (DPAM). The authors report VisDrone experiments showing accuracy gains on DETR, RT-DETR, SSD, YOLO11s, and Faster R-CNN, plus an ablation study on YOLO11s. The abstract additionally claims YOLO26m, UAVDT, and CARPK results that do not appear in the full text.

Significance. If the reported gains are reproducible, the modular, detector-agnostic design could be a practical addition to small-object detection pipelines, especially because the ablation shows incremental contributions from each component. However, the paper's evidentiary value is currently limited: the headline quantitative claims in the abstract are unsupported by any experiment in the main text, and the formal definitions in Section III contain undefined symbols and dimensional inconsistencies. The manuscript does not provide code for verification beyond a GitHub link, and no machine-checked proofs or reproducible evaluation harness are included. The central claim of consistent improvement across detectors and benchmarks therefore cannot be assessed from the written evidence.

major comments (5)
  1. [Abstract vs. Section IV] The abstract reports that on a 'controlled VisDrone setting,' YOLO26m + MGDFIS improves AP50:95 from 25.7 to 30.2 and AP50 from 37.2 to 44.2 at 96.1 GFLOPs, and additionally reports UAVDT AP50 38.9 / AP50:95 21.9 and CARPK AP50 97.4. None of these experiments appear in Section IV: Tables I and II cover DETR, RT-DETR, SSD, YOLO11s, Faster R-CNN, and YOLOv5s through YOLOv11s, while Table III is the YOLO11s ablation. There is no YOLO26m configuration, no UAVDT table, no CARPK table, no dataset split, and no hyperparameters for those results. These claims are the headline quantitative evidence for 'significant enhancement' and must either be supported with full experimental details or removed from the abstract.
  2. [Section III-A, Eq. (7)] Equation (7) defines FCom = TSSA(x) = |Reshape(−FStos ∗ π ∗ Attn)|_{LN}, but the symbols FStos and π are never defined. The preceding equations define FSoms and Π, but no relation between (FStos, π) and (FSoms, Π) is given. This makes the formal definition of the core attention module incomplete and prevents a reader from verifying the architecture or reproducing the implementation.
  3. [Section IV-B1 and Table I] Table I reports that Faster R-CNN with MGDFIS increases AP50 from 40.7% to 53.2% (a 12.5 percentage-point gain) while mAP increases only from 31.4% to 33.4%. Such a large AP50 gain with a small mAP gain is atypical and suggests a protocol or implementation detail not captured in the paper. The text should report per-class AP, the exact training/validation split, inference settings (NMS, resolution, multi-scale test), and clarify whether MGDFIS modifies only the neck or also changes the training schedule. In addition, Section IV-A states all models are trained with AdamW, batch size 64, and 400 epochs, whereas Section IV-B1 says experiments were conducted 'without altering their original training configurations'; these statements need to be reconciled.
  4. [Section IV-B1] The sentence 'Faster R-CNN also exhibits the most significant gains in AP50 and AP75, improving by 12.5% and 14.8% respectively, approaching the performance of DN-FPN' references DN-FPN without a citation and without a comparison in any table. Either provide the citation and the corresponding quantitative comparison, or remove the reference to DN-FPN.
  5. [Section III-A, Eqs. (4), (5), and (8)] Equation (4) sums FSoms over the D dimension and multiplies by E_{H×H}, but FSoms is stated to be in R^{B×H×N×D}; the dimensions are inconsistent. Equation (5) uses Π both as a probability distribution and as a summation variable, and the expression is ambiguous. Equation (8) multiplies a feature x ∈ R^{C×H×W} by E_{C×C} × 1e−6 × E_{C×C}, which is not a valid tensor operation on the stated shapes and is undefined. These formal issues make the method difficult to understand and impossible to verify as written.
minor comments (5)
  1. [Throughout] There are several typos and formatting issues: 'UA V imagery' in the abstract, 'Intergation' in Section III, 'The code are available' in the abstract, and two inconsistent GitHub URLs (github.com/Bai-Xuecheng/MGDFIS in the abstract and github.com/JackBaixue/MGDFIS in the full text).
  2. [Section IV] The terms AP and mAP are used interchangeably in the text and tables; the authors should define whether the reported 'AP' column denotes AP50:95 or something else. Table I reports only AP and AP50, while the text discusses AP75 gains; the table should include AP75 and ideally computational cost (GFLOPs) for the compared models.
  3. [Algorithm 1] In Algorithm 1, lines 5–6 assign Fagg = Ffuse inside the loop, but Ffuse is first computed later in line 10; this makes the pseudocode non-executable and should be corrected.
  4. [Section III-A2, Eq. (16)] The symbol FSEEF in Eq. (16) appears to be a typo for FSEFF defined in Eq. (15); please harmonize the notation.
  5. [Section IV-B3] The ablation study in Table III reports GFLOPs but not inference time or parameter counts; since the abstract claims computational efficiency, adding these metrics for each configuration would strengthen the efficiency claim.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the paper reports a modular architecture evaluated on held-out benchmarks, with no fitted parameter or self-citation chain that reduces the central claim to its own inputs.

full rationale

The central claim is that the proposed MGDFIS architecture improves small-object detection accuracy on VisDrone and other benchmarks. This claim is supported by comparative experiments and ablations in Tables I-III, where the method is integrated into standard detectors and evaluated on held-out test data. The module definitions, despite notational issues such as undefined FStos and pi in Eq. (7), do not define a target metric in terms of an input or fit a parameter to the same data that is later 'predicted'. The method is an architectural modification trained with standard supervised learning, and the reported gains are empirical outcomes, not consequences of a self-referential definition. No load-bearing uniqueness theorem, ansatz, or empirical law is imported solely from the authors' own prior work. Issues of missing experimental tables and ambiguous formal definitions are correctness and reproducibility concerns, not circularity. Therefore, the derivation chain is self-contained with respect to circularity, and no specific reduction of a prediction to an input can be exhibited.

Assumptions & free parameters 2 free parameters · 3 assumptions · 0 invented entities

No new physical entities are introduced. The reported modules are architectural components built from published ingredients. The free parameters are small numerical constants in the normalization and residual scaling.

free parameters (2)
  • epsilon in Dots normalization (Eq. 5) = 1e-8
    Ad hoc small constant added to prevent division by zero in weighted second-order statistics; not justified by theory.
  • epsilon in XMona scaling (Eq. 8) = 1e-6
    Hand-chosen small factor in the XMona residual branch; no derivation given.
assumptions (3)
  • domain assumption VisDrone is a representative benchmark for UAV small-object detection.
    All conclusions are drawn from this dataset; the paper treats it as the standard evaluation.
  • domain assumption The referenced modules TSSA, DyT, Mona, and SEFF behave as published.
    The paper builds on these methods without re-deriving them; if they differ in the context of MGDFIS, the attributed gains may be wrong.
  • domain assumption The baseline detectors are trained and evaluated with correct, consistent settings.
    The paper states that all baselines use the same hyperparameters, but it does not provide sufficient logs to verify this.

how reviews work

0 comments
Cite this review

Pith. "Pith review of MGDFIS: Multi-scale Global-detail Feature Integration Strategy for Small Object Detection." pith.science (2026). https://pith.science/paper/XYEFKJF5

@misc{pith2026250612697,
  author       = {Pith},
  title        = {Pith review of: MGDFIS: Multi-scale Global-detail Feature Integration Strategy for Small Object Detection},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/XYEFKJF5}},
  note         = {Machine review of arXiv:2506.12697}
}
read the original abstract

Small-object detection in Unmanned Aerial Vehicle (UAV) imagery requires preserving weak local evidence while using broader context to separate tiny foreground targets from cluttered backgrounds. Existing multi-scale fusion methods improve feature aggregation, but they often add computation or blur fine details during repeated cross-scale fusion. The central challenge is to balance low-SNR target preservation, clutter suppression, and efficient cross-scale context exchange. To address this challenge, we propose the Multi-scale Global-detail Feature Integration Strategy (MGDFIS), a neck-level feature-fusion strategy that couples global context exchange, local-detail recovery, and pixel-level foreground-background recalibration. MGDFIS integrates three coordinated modules: FusionLock-TSS Attention for stabilizing spectral-spatial responses, Global-detail Integration for combining long-range mixing with local detail capture, and Dynamic Pixel Attention for reweighting compact foreground regions. On the controlled VisDrone setting, YOLO26m + MGDFIS improves AP50:95 from 25.7 to 30.2 and AP50 from 37.2 to 44.2 over the YOLO26m baseline, with 96.1 GFLOPs. Additional dataset-specific evaluations report 38.9 AP50 and 21.9 AP50:95 on UAVDT and 97.4 AP50 on CARPK. The code is available at: https://github.com/Bai-Xuecheng/MGDFIS.

Figures

Figures reproduced from arXiv: 2506.12697 by the authors.

Figure 1
Figure 1. Comparison of traditional feature fusion strategies and our method. Conventional feature fusion approaches suffer from [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. FusionLock-TSS Attention integrates DAFF and SERR in series: DAFF filters critical features via DynamicTanh and TSSA attention, fusing spatial contexts with MonaOp’s depthwise-separable convolutions. SERR enhances global dependencies through dilated convolutions and FFT-weighting in FFN, then refines details via MonaOp. Residual skip-connections aggregate outputs to generate high-SNR multi-scale global-local fused f… view at source ↗
Figure 3
Figure 3. The figure illustrates our proposed MGDFIS: subfigure (a) shows the overall framework schematic, subfigure (b) depicts the GMM structure within the GDIM, and subfigure (c) depicts the DMM structure within the GDIM. and and the refined feature Fˆ, obtained from the input and GDIM respectively, are fed into DPAM to generate an attention weight map AM ap ∈ [0, 1]B×C×H×W , which encodes both the coarse-detail informatio… view at source ↗
Figures from the paper (2 more)
Figure 4
Figure 4. Figure 4: Structure of the Dynamic Pixel Attention Module especially well-suited for tasks requiring meticulous feature processing. IV. EXPERIMENTS A. Datasets and Evaluation Measures • Dataset: We conduct our experiments on the VisDrone 2019 [19] drone dataset to validate the e…
Figure 5
Figure 5. Figure 5: Visualization of detection results and heatmaps on the VisDrone dataset. The highlighted areas represent the key focus [PITH_FULL_IMAGE:figures/full_fig_p008_5.png]

Discussion (0). Continue with ORCID to comment.

Forward citations

Cited by 1 Pith paper

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

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

    cs.CV 2026-03 conditional novelty 4.0 of 10

    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

35 extracted references · 30 canonical work pages · cited by 1 Pith paper

  1. [1]

    Remote sensing analysis of agricultural drone,

    S. Meivel and S. Maheswari, “Remote sensing analysis of agricultural drone,” Journal of the Indian Society of Remote Sensing , vol. 49, no. 3, pp. 689–701, 2021

    work page 2021

  2. [2]

    Efficient drone-based rare plant monitoring using a species distribution model and ai-based object detection,

    W. Reckling, H. Mitasova, K. Wegmann, G. Kauffman, and R. Reid, “Efficient drone-based rare plant monitoring using a species distribution model and ai-based object detection,” Drones, vol. 5, no. 4, p. 110, 2021

    work page 2021

  3. [3]

    A Survey of 3D Reconstruction with Event Cameras

    C. Xu, H. Zhou, L. Chen, H. Chen, Y . Zhou, V . Chung, and Q. Qu, “A survey of 3d reconstruction with event cameras,” arXiv preprint arXiv:2505.08438, 2025

    work page Pith review arXiv 2025

  4. [4]

    Autonomous monitoring, analysis, and countering of air pollution using environmental drones,

    G. Rohi, O. Ejofodomi, and G. Ofualagba, “Autonomous monitoring, analysis, and countering of air pollution using environmental drones,” Heliyon, vol. 6, no. 1, 2020

    work page 2020

  5. [5]

    The use of drones and au- tonomous vehicles in logistics and delivery,

    I. Nurgaliev, Y . Eskander, and K. Lis, “The use of drones and au- tonomous vehicles in logistics and delivery,” Logistics and Transport , vol. 57, 2023

    work page 2023

  6. [6]

    Using the unmanned aerial vehicle delivery decision tool to consider transporting medical supplies via drone,

    M. Eichleay, E. Evens, K. Stankevitz, and C. Parker, “Using the unmanned aerial vehicle delivery decision tool to consider transporting medical supplies via drone,” Global Health: Science and Practice, vol. 7, no. 4, pp. 500–506, 2019

    work page 2019

  7. [7]

    Tiny object detection with context enhancement and feature purification,

    J. Xiao, H. Guo, J. Zhou, T. Zhao, Q. Yu, Y . Chen, and Z. Wang, “Tiny object detection with context enhancement and feature purification,” Expert Systems with Applications , vol. 211, p. 118665, 2023

    work page 2023

  8. [8]

    A survey and performance evaluation of deep learning methods for small object detection,

    Y . Liu, P. Sun, N. Wergeles, and Y . Shang, “A survey and performance evaluation of deep learning methods for small object detection,” Expert Systems with Applications , vol. 172, p. 114602, 2021

    work page 2021

Show all 35 references
  1. [9]

    To- wards large-scale small object detection: Survey and benchmarks,

    G. Cheng, X. Yuan, X. Yao, K. Yan, Q. Zeng, X. Xie, and J. Han, “To- wards large-scale small object detection: Survey and benchmarks,” IEEE Transactions on Pattern Analysis and Machine Intelligence , vol. 45, no. 11, pp. 13 467–13 488, 2023

    2023

  2. [10]

    Methods for small, weak object detection in optical high-resolution remote sensing images: A survey of advances and challenges,

    W. Han, J. Chen, L. Wang, R. Feng, F. Li, L. Wu, T. Tian, and J. Yan, “Methods for small, weak object detection in optical high-resolution remote sensing images: A survey of advances and challenges,” IEEE Geoscience and Remote Sensing Magazine, vol. 9, no. 4, pp. 8–34, 2021

    2021

  3. [11]

    Towards end-to- end neuromorphic voxel-based 3d object reconstruction without physical priors,

    C. Xu, L. Chen, H. Chen, V . Chung, and Q. Qu, “Towards end-to- end neuromorphic voxel-based 3d object reconstruction without physical priors,” arXiv preprint arXiv:2501.00741 , 2025

    2025

  4. [12]

    Effective fusion factor in fpn for tiny object detection,

    Y . Gong, X. Yu, Y . Ding, X. Peng, J. Zhao, and Z. Han, “Effective fusion factor in fpn for tiny object detection,” in Proceedings of the IEEE/CVF winter conference on applications of computer vision , 2021, pp. 1160–1168

    2021

  5. [13]

    Face mask wearing detection algorithm based on improved yolo-v4,

    J. Yu and W. Zhang, “Face mask wearing detection algorithm based on improved yolo-v4,” Sensors, vol. 21, no. 9, p. 3263, 2021

    2021

  6. [14]

    A multi-scale small object detection algorithm sma-yolo for uav remote sensing images,

    S. Zhou, H. Zhou, and L. Qian, “A multi-scale small object detection algorithm sma-yolo for uav remote sensing images,” Scientific Reports, vol. 15, no. 1, p. 9255, 2025

    2025

  7. [15]

    Dcn-yolo: A small-object detection paradigm for remote sensing imagery leveraging dilated convolutional networks,

    M. Xie, Q. Tang, Y . Tian, X. Feng, H. Shi, and W. Hao, “Dcn-yolo: A small-object detection paradigm for remote sensing imagery leveraging dilated convolutional networks,” Sensors (Basel, Switzerland) , vol. 25, no. 7, p. 2241, 2025

    2025

  8. [16]

    Gcl-yolo: A ghostconv-based lightweight yolo network for uav small object detec- tion,

    J. Cao, W. Bao, H. Shang, M. Yuan, and Q. Cheng, “Gcl-yolo: A ghostconv-based lightweight yolo network for uav small object detec- tion,” Remote Sensing, vol. 15, no. 20, p. 4932, 2023

    2023

  9. [17]

    Yolo-tla: an efficient and lightweight small object detection model based on yolov5,

    C.-L. Ji, T. Yu, P. Gao, F. Wang, and R.-Y . Yuan, “Yolo-tla: an efficient and lightweight small object detection model based on yolov5,” Journal of Real-Time Image Processing , vol. 21, no. 4, p. 141, 2024

    2024

  10. [18]

    An anchor-free network for increasing attention to small objects in high resolution remote sensing images,

    H. Zhou, W. Guo, and Q. Zhao, “An anchor-free network for increasing attention to small objects in high resolution remote sensing images,” Applied Sciences, vol. 13, no. 4, p. 2073, 2023

    2023

  11. [19]

    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 . Zhang et al. , “Visdrone-det2019: The vision meets drone object detection in image challenge results,” in Proceedings of the IEEE/CVF international conference on computer vision workshops , 2019, pp. 0–0

    2019

  12. [20]

    Starting from the structure: A review of small object detection based on deep learning,

    Z. Xiuling, W. Huijuan, S. Yu, C. Gang, Z. Suhua, and Y . Quanbo, “Starting from the structure: A review of small object detection based on deep learning,” Image and Vision Computing , p. 105054, 2024

    2024

  13. [21]

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

    2022

  14. [22]

    Depth-first random forests with improved grassberger entropy for small object detection,

    J. Ma, Q. Pan, and Y . Guo, “Depth-first random forests with improved grassberger entropy for small object detection,” Engineering Applica- tions of Artificial Intelligence , vol. 114, p. 105138, 2022

    2022

  15. [23]

    Banet: Small and multi- object detection with a bidirectional attention network for traffic scenes,

    S.-y. Wang, Z. Qu, C.-j. Li, and L.-y. Gao, “Banet: Small and multi- object detection with a bidirectional attention network for traffic scenes,” Engineering Applications of Artificial Intelligence , vol. 117, p. 105504, 2023

    2023

  16. [24]

    Extended feature pyramid network for small object detection,

    C. Deng, M. Wang, L. Liu, Y . Liu, and Y . Jiang, “Extended feature pyramid network for small object detection,” IEEE Transactions on Multimedia, vol. 24, pp. 1968–1979, 2021

    1968

  17. [25]

    Enhanced semantic feature pyramid network for small object detection,

    Y . Chen, X. Zhu, Y . Li, Y . Wei, and L. Ye, “Enhanced semantic feature pyramid network for small object detection,” Signal Processing: Image Communication, vol. 113, p. 116919, 2023

    2023

  18. [26]

    Dense and small object detection in uav-vision based on a global-local feature enhanced network,

    T. Ye, W. Qin, Y . Li, S. Wang, J. Zhang, and Z. Zhao, “Dense and small object detection in uav-vision based on a global-local feature enhanced network,” IEEE Transactions on Instrumentation and Measurement , vol. 71, pp. 1–13, 2022

    2022

  19. [27]

    Token statistics transformer: Linear-time attention via variational rate reduction,

    Z. Wu, T. Ding, Y . Lu, D. Pai, J. Zhang, W. Wang, Y . Yu, Y . Ma, and B. D. Haeffele, “Token statistics transformer: Linear-time attention via variational rate reduction,” arXiv preprint arXiv:2412.17810 , 2024

    2024 arXiv

  20. [28]

    Transformers without normalization,

    J. Zhu, X. Chen, K. He, Y . LeCun, and Z. Liu, “Transformers without normalization,” in Proceedings of the Computer Vision and Pattern Recognition Conference, 2025, pp. 14 901–14 911

    2025

  21. [29]

    A hybrid transformer-mamba network for single image deraining,

    S. Sun, W. Ren, J. Zhou, J. Gan, R. Wang, and X. Cao, “A hybrid transformer-mamba network for single image deraining,” arXiv preprint arXiv:2409.00410, 2024

    2024 arXiv

  22. [30]

    5%¿ 100%: Breaking performance shackles of full fine-tuning on visual recognition tasks,

    D. Yin, L. Hu, B. Li, Y . Zhang, and X. Yang, “5%¿ 100%: Breaking performance shackles of full fine-tuning on visual recognition tasks,” in Proceedings of the Computer Vision and Pattern Recognition Confer- ence, 2025, pp. 20 071–20 081

    2025

  23. [31]

    Efficientdet: Scalable and efficient object detection,

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

    2020

  24. [32]

    Ultralytics YOLO,

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

    2023

  25. [33]

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

    A. Wang, H. Chen, L. Liu, K. Chen, Z. Lin, J. Han et al. , “Yolov10: Real-time end-to-end object detection,” Advances in Neural Information Processing Systems, vol. 37, pp. 107 984–108 011, 2024

    2024

  26. [34]

    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,” in Euro- pean conference on computer vision . Springer, 2024, pp. 1–21

    2024

  27. [35]

    Yolov6: A single-stage object detection framework for industrial applications,

    C. Li, L. Li, H. Jiang, K. Weng, Y . Geng, L. Li, Z. Ke, Q. Li, M. Cheng, W. Nie et al. , “Yolov6: A single-stage object detection framework for industrial applications,” arXiv preprint arXiv:2209.02976 , 2022

    2022 arXiv

Pith tools

Reviewed August 7, 2026 · model on record in the stance chip above.