Pith. sign in

REVIEW 4 major objections 5 minor 47 references

Attention from Above: A Multimodal Model for Drone-Based Object Localization

T0 review · 4 major / 5 minor · reviewed 2026-08-01 · deepseek-v4-flash

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

Pith's one-line read Replacing YOLO-World's C2f blocks with attention-based A2C2f layers raises VisDrone mAP@0.5 from 32.5 to 35.2.

desk verdict A plausible YOLO-World backbone swap for VisDrone, but the paper's own numbers contradict each other and the missing experimental controls make the headline gain unverifiable. read the letter →

arxiv 2607.17669 v1 pith:UQ4RUKEO submitted 2026-07-20 cs.CV

Hyun-Ki Jung This is my paper
classification cs.CV
keywords objectdetectiondroneimageryYOLO-WorldA2C2fattentionmechanismsmalltext-guidedVis
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

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

The reading

The paper tries to establish that a targeted architectural swap can make text-guided object detection from drone footage better at small objects. It takes YOLO-World, which fuses CLIP text embeddings with YOLOv8-style visual features, and replaces the backbone's C2f building blocks with A2C2f blocks that add a lightweight area-attention mechanism. On the VisDrone benchmark, the modified model reports gains over YOLO-World on every headline metric: precision rises from 43.0 to 45.1, recall from 32.8 to 35.0, F1 from 37.2 to 39.4, mAP@0.5 from 32.5 to 35.2, and mAP@0.5-0.95 from 18.5 to 19.9. If the comparison is reproducible, the recipe is attractive because it changes only the backbone and leaves the text encoder and the rest of the pipeline intact.

What carries the argument

A2C2f layer: an attention-augmented successor to YOLOv8's C2f block. It embeds an Area Attention (A2) module that recalibrates feature responses through adaptive attention aggregation, and it parallelizes ABlock layers that combine multi-head MLP blocks with localized area-attention; the modified backbone also uses C2PSA for spatial attention. Its job is to emphasize informative regions and suppress background without the quadratic cost of global attention, making it a lightweight swap for the original C2f blocks.

What would settle it

Run both models on VisDrone with an identical fixed prompt list, at least five seeds, and the same hyperparameters; if the mean mAP@0.5 difference between the proposed model and YOLO-World is near zero (say under 1 point) or changes sign, the central claim fails. A cheaper check: reconcile Section 5's mAP@0.5 of 30.7% with Table 2's 35.2%—they cannot both be correct for the same run.

Watch

Extended reading notes

Core claim

The central claim is that the C2f layer—the residual building block in YOLOv8's backbone—is a bottleneck for small-object representation, and replacing it with A2C2f layers lets the model concentrate on informative local regions while keeping a large receptive field. The paper reports that this single change, plus a C2PSA spatial-attention block in the modified backbone, lifts every standard detection metric on VisDrone relative to the original YOLO-World, with the largest reported gain a 2.7-point increase in mAP@0.5. The model also scores higher than YOLOv9, YOLOv10, YOLOv11, YOLOv12, and YOLOv26 on the same evaluation.

Load-bearing premise

The comparison assumes the YOLO-World baseline and the proposed model were trained and evaluated under identical, fair conditions—same prompt strings, seeds, optimizer settings, and epochs—so that the 2.7-point mAP@0.5 gap is caused by the A2C2f swap rather than by configuration differences or random variation, especially since the experiments used default YOLOv8n settings with no seeds or code provided.

Editorial extensions

If this is right

  • If the reported comparison is reproducible, any YOLO-World deployment for drone imagery can get a consistent accuracy bump by swapping the backbone blocks, with no change to the text encoder, neck, or head.
  • The design suggests attention at moderate cost is enough to help small, boundary-defined classes; on VisDrone the biggest per-class mAP@0.5 is cars (77.4), while bicycles and awning-tricycles remain the hardest.
  • The gains appear across all standard detection metrics, not just one threshold, which is consistent with a feature-representation improvement rather than a threshold artifact.
  • Because the prompts in the experiments are free-text commands, the approach also demonstrates that open-vocabulary text-guided drone detection can improve without retraining the language side.

Reading between the lines

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

  • The paper leaves untested whether the attention swap transfers to other drone datasets or to other YOLOv8-derived open-vocabulary detectors; testing on aerial scenes with different object densities would show how general the 2.7-point gain is.
  • The per-class results suggest the benefit is uneven: cars and vans improve strongly while bicycles and tricycles remain hard, so a confusion-matrix analysis could reveal whether the gain comes from suppressing false positives on background or from recovering missed small objects.
  • A contradiction inside the paper needs resolving: Section 5 reports mAP@0.5 of 30.7% and a 0.3-point gain over YOLO-World, whereas Table 2 reports 35.2% and a 2.7-point gain; both cannot describe the same model, so the effect size is currently uncertain.
  • The prompt-based comparisons in Table 4 are qualitative images rather than scored outputs; converting them into a benchmark with fixed prompts, multiple seeds, and quantitative metrics would separate the backbone's contribution from prompt sensitivity.
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

4 major / 5 minor

Summary. The paper proposes a multimodal object detection model for drone imagery, built on YOLO-World by replacing the C2f layers of the YOLOv8 backbone with 'A2C2f' attention-based layers. The authors report experiments on the VisDrone dataset claiming consistent improvements over the original YOLO-World in precision, recall, F1, mAP@0.5, and mAP@0.5-0.95, as well as comparisons with YOLOv9-v26. The central claim is that the A2C2f swap yields a simple, effective gain for small-object detection in text-guided drone detection.

Significance. If the reported gains are reliable, the proposed A2C2f backbone replacement could be a low-cost improvement to YOLO-World for drone-view small-object detection, and the paper would offer a useful empirical datapoint. The paper also provides class-wise results (Table 3) and prompt-based qualitative comparisons in the text. However, the manuscript is currently not reproducible from the information given: no code, no seeds, no prompt strings, and no detailed training configurations. The internal contradiction between the Abstract/Table 2 and the Conclusions further undermines confidence. The strengths—clear architecture description and use of a standard benchmark—do not compensate for the missing experimental rigor needed to substantiate the central improvement claim.

major comments (4)
  1. [Section 5 (Conclusions)] The Conclusions report that the proposed model achieved mAP@0.5 of 30.7% with a 0.3% improvement and mAP@0.5-0.95 of 19.9% with a 0.4% improvement. This directly contradicts the Abstract and Table 2, which report mAP@0.5 of 35.2% (a 2.7% gain) and mAP@0.5-0.95 of 19.9% (a 1.4% gain). Both cannot be correct. Since the manuscript's central claim rests on the magnitude of improvement, this discrepancy is a load-bearing error that must be resolved before any assessment of the method's effectiveness can be made.
  2. [Section 4.1 / Table 2] The comparison between the proposed model and YOLO-World is not properly controlled. The text states only that experiments used the 'default configuration of YOLOv8n' with 100 epochs and 640×640 input; no text prompts, random seed, optimizer, batch size, learning rate, or evaluation protocol are given. For a text-conditioned detector like YOLO-World, the prompt vocabulary directly determines which classes are decoded; if the baseline and proposed model were given different prompts, the 2.7-point mAP gap is confounded. Additionally, all numbers are single runs with no error bars or multiple seeds; a 2.7-point difference on a nano-scale model could fall within run-to-run noise. The authors must specify the exact prompts, training details, and repeat experiments with multiple seeds (or provide a statistical analysis) to isolate the effect of the A2C2f block.
  3. [Table 4] Table 4, titled 'Comparison of Object Detection Results by Command According to Prompt Text Input,' is empty in the submitted manuscript. The rows for YOLO-World and Proposed Model contain no values or images. The text in Section 4.2 states that comparative results are presented in Table 4, but no data are shown. This is a missing result that directly supports the prompt-dependent behavior claimed in the same section. The table must be populated with actual detection outputs or quantitative metrics before the claim can be assessed.
  4. [Table 2 / Section 4.2] The comparisons against YOLOv9, YOLOv10, YOLOv11, YOLOv12, and YOLOv26 lack any training or evaluation details for these models. It is not stated whether they were trained on the same train split with the same epochs, input resolution, optimizer, or any hyperparameters, nor whether they use the same test-time settings. Without this information, the reported superiority over these models (e.g., mAP@0.5 of 35.2% vs. 33.0% for YOLOv9) cannot be interpreted. The authors should either provide full experimental settings for all baselines or remove these comparisons and focus solely on the controlled YOLO-World comparison.
minor comments (5)
  1. [Abstract / Section 4.2] The phrase 'substantial improvement' is used for gains of 2.1-2.7 points. Given the lack of statistical significance testing, this wording is too strong; 'reported improvement' would be more appropriate.
  2. [Section 2.2 / Reference [45]] The C2PSA module is attributed to the CSPNet reference [45], but C2PSA is a YOLOv8-specific module with a spatial attention mechanism. The cited reference does not appear to introduce C2PSA. The authors should cite the correct source (e.g., the Ultralytics implementation or the original paper introducing the spatial attention mechanism).
  3. [Table 2] The 'YOLOv26' row appears to be based on a future or nonstandard model; the reference list does not include a YOLOv26 paper. Please clarify the model's origin or remove it.
  4. [Section 4.2 / Figure 5] The text refers to 'the graph on the left side of Figure 5(a)' and 'the graph on the right side of Figure 5(b),' but Figure 5 appears to be a single image with two panels. The caption should be clarified, and the panels should be labeled more explicitly.
  5. [Section 3.3, Eq. (4)] Equation (4) shows the integral from 0 to 1 of P(r) dr, but the standard AP formulation integrates over recall. The notation should be corrected or explained to avoid confusion.

Circularity Check

0 steps flagged · score 2.0 of 10

No derivation-level circularity; the claimed gain is an empirical architecture-swap comparison against an external baseline. The only self-citation is minor and not load-bearing.

full rationale

The paper's central claim is that replacing the C2f layers in the YOLO-World backbone with A2C2f attention layers improves VisDrone detection metrics. This is an empirical comparison, not a derived identity: the A2C2f and C2PSA modules are explicitly attributed to external prior work ([44], [45]), and the baseline YOLO-World is external ([43]). No equation in the paper links the architecture definition to the reported metric values, and no fitted parameter is renamed as a prediction. The only self-citation is [46] in the Conclusion, 'building upon previous research [46]', which is a continuity statement and does not carry a load-bearing premise, forbid alternatives, or supply a uniqueness claim. I therefore find no circular derivation. Two non-circular concerns are flagged: (i) Section 5 reports 'mAP@0.5 and mAP@0.5–0.95 scores of 30.7% and 19.9%, respectively. These results correspond to improvements of 0.3% and 0.4% over the original YOLO-World model', which contradicts the Abstract and Table 2 values of 35.2% and 2.7%; this is an internal inconsistency that undermines reproducibility but is not a circular step. (ii) The experimental protocol does not specify prompt strings, seeds, optimizer settings, or multiple runs, so the comparison is not fully controlled; this is a validity/reproducibility risk, not circularity. Score 2 reflects only the minor, non-load-bearing self-citation.

Assumptions & free parameters 4 free parameters · 4 assumptions · 0 invented entities

No genuinely new entity or theoretical construct is introduced; A2C2f combines the cited A2 attention module with C2f/C2PSA structures. The empirical claim rests on borrowed modules and training choices that are only partially reported.

free parameters (4)
  • YOLOv8n scale (depth multiple, width multiple) = 0.33, 0.25
    Chosen from the default YOLOv8n configuration in Section 4.1, not justified for VisDrone; all comparisons use this capacity and the performance delta may depend on it.
  • Training schedule (epochs, input resolution) = 100 epochs, 640x640
    Hand-set in Section 4.1 with no ablation or sensitivity analysis; the reported gains are conditional on this schedule.
  • Text prompts for VisDrone classes
    YOLO-World is text-guided, but the exact prompt strings and possible negative prompts are not specified, and these can materially affect open-vocabulary detection outputs.
  • Unspecified optimizer, batch size, learning rate
    Not reported; needed to reproduce the baseline-versus-modified comparison.
assumptions (4)
  • domain assumption Pretrained CLIP text encoder and YOLO-World/RepVL-PAN modules transfer to VisDrone without domain-specific adaptation beyond fine-tuning the visual backbone.
    Section 3.1 freezes CLIP and uses pretrained YOLO-World components; no analysis of domain shift to aerial imagery is provided.
  • ad hoc to paper The A2 attention module from [44] and C2PSA structures remain effective when inserted into YOLO-World's YOLOv8 backbone and do not require architectural changes to RepVL-PAN.
    Section 3.2 states the replacement, but no compatibility proof or component ablation is provided.
  • domain assumption VisDrone ground-truth labels and standard COCO-style mAP evaluation are correct and applicable at 640x640 resolution without resolution-specific recalibration.
    Section 4.1 sets the input resolution to 640; VisDrone objects are small, and no annotation-quality or resolution analysis is given.
  • standard math Standard precision, recall, F1, AP, and mAP definitions accurately summarize detection quality.
    Equations (1)-(5) define standard metrics that are used without proof.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Attention from Above: A Multimodal Model for Drone-Based Object Localization." pith.science (2026). https://pith.science/paper/UQ4RUKEO

@misc{pith2026260717669,
  author       = {Pith},
  title        = {Pith review of: Attention from Above: A Multimodal Model for Drone-Based Object Localization},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/UQ4RUKEO}},
  note         = {Machine review of arXiv:2607.17669}
}
read the original abstract

Drone-based object detection technology has advanced rapidly, becoming increasingly sophisticated and efficient. Recently, research trends have expanded beyond the detection of predefined objects toward the identification of specified target objects. For example, desired targets can be specified through textual prompts, enabling accurate detection of objects of interest. To address this demand, this paper proposes an efficient multimodal-based object detection model aimed at improving small object detection performance. The proposed method is built upon the YOLO-World framework and replaces the C2f layers used in the YOLOv8 backbone with attention-based A2C2f layers. This modification enables more precise representation of local features, particularly for small objects or objects with well-defined boundaries. In addition, the incorporation of attention mechanisms and parallel processing structures significantly enhances the model's computational accuracy. Comparative experiments conducted on the VisDrone dataset demonstrate that the proposed model outperforms the original YOLO-World model. Specifically, precision increases from 43.0% to 45.1%, recall from 32.8% to 35.0%, the F1 score from 37.2% to 39.4%, mAP@0.5 from 32.5% to 35.2%, and mAP@0.5-0.95 from 18.5% to 19.9%, confirming a substantial improvement in detection accuracy. These results verify that the proposed approach provides an effective and highly accurate solution for object detection in drone-based image and video application environments.

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

47 extracted references · 12 canonical work pages

  1. [1]

    To achieve this goal, extensive experiments were conducted using the VisDrone dataset, which consists of drone- captured images collected under diverse environmental conditions

    The primary objective of this paper is to propose a text-guided object detection model capable of accurately detecting small objects. To achieve this goal, extensive experiments were conducted using the VisDrone dataset, which consists of drone- captured images collected under diverse environmental conditions. Based on these experiments, a multimodal obje...

  2. [2]

    car” and “truck

    This study proposes a novel network architecture in which the C2f layers of the original YOLO-World backbone are replaced with attention-based A2C2f layers. As a result, performance improvements were observed across multiple evaluation metrics, including precision, recall, F1 score, mAP@0.5, and mAP@0.5–0.95. In particular, the mAP@0.5 increased from 32.5...

    1975

  3. [3]

    From toys to tools: The co-evolution of technological and entrepreneurial developments in the drone industry,

    F. Giones, B. Alexander, "From toys to tools: The co-evolution of technological and entrepreneurial developments in the drone industry," Business Horizons, vol. 60, no. 6, pp. 875-884, 2017. https://doi.org/10.1016/j.bushor.2017.08.001

    work page doi:10.1016/j.bushor.2017.08.001 2017

  4. [4]

    Edge Computing, Emerging Trends, and Drone Technology for the Business Industry,

    I. A. Shah, A. Manzoor, "Edge Computing, Emerging Trends, and Drone Technology for the Business Industry," Computer Vision and Edge Computing Technologies for the Drone Industry, pp. 89-104, 2025. https://doi.org/10.4018/979-8-3693-8497-8.ch004

    work page doi:10.4018/979-8-3693-8497-8.ch004 2025

  5. [5]

    Use of AI applications for the drone industry,

    I. A. Shah, et al, "Use of AI applications for the drone industry," Cybersecurity Issues and Challenges in the Drone Industry, pp. 27-41, 2024. https://doi.org/10.4018/979-8-3693-0774- 8.ch002

    work page doi:10.4018/979-8-3693-0774- 2024

  6. [6]

    UFPMP-Det:Toward Accurate and Efficient Object Detection on Drone Imagery,

    Y. Huang, et al, "UFPMP-Det:Toward Accurate and Efficient Object Detection on Drone Imagery," In Proceedings of the AAAI conference on artificial intelligence, vol. 36, no. 1, pp. 1026- 1033, 2022. https://doi.org/10.1609/aaai.v36i1.19986

    work page doi:10.1609/aaai.v36i1.19986 2022

  7. [9]

    DR-YOLO: An improved multi-scale small object detection model for drone aerial photography scenes based on YOLOv7,

    H. Bi, et al, "DR-YOLO: An improved multi-scale small object detection model for drone aerial photography scenes based on YOLOv7," Digital Signal Processing, 2025. https://doi.org/10.1016/j.dsp.2025.105265

    arXiv 2025

  8. [10]

    DHANet: Dual-Stream Hierarchical Interaction Networks for Multimodal Drone Object Detection,

    X. Wu, et al, "DHANet: Dual-Stream Hierarchical Interaction Networks for Multimodal Drone Object Detection," IEEE Transactions on Geoscience and Remote Sensing, 2025. https://doi.org/10.1109/TGRS.2025.3578675

    arXiv 2025

Show all 47 references
  1. [11]

    HiCAL: Hierarchical Consistency-Based Active Learning for Drone-View Object Detection,

    X. Zhang, et al, "HiCAL: Hierarchical Consistency-Based Active Learning for Drone-View Object Detection," IEEE Transactions on Geoscience and Remote Sensing, 2025. https://doi.org/10.1109/TGRS.2025.3576980

    2025

  2. [12]

    Automated Detection and Monitoring of Ground-Nesting Bee Nests Using Drone Imagery and Deep Learning,

    P. Tschanz, et al, "Automated Detection and Monitoring of Ground-Nesting Bee Nests Using Drone Imagery and Deep Learning," Ecology and Evolution, vol. 16, no. 1, 2026. https://doi.org/10.1002/ece3.72856

    2026 doi

  3. [13]

    SPVD-DETR: A novel real-time end-to-end object detector of sweetpotato virus disease from unmanned aerial vehicle ortho imagery,

    F. Zeng, et al, "SPVD-DETR: A novel real-time end-to-end object detector of sweetpotato virus disease from unmanned aerial vehicle ortho imagery," Computers and Electronics in Agriculture, 2026. https://doi.org/10.1016/j.compag.2025.111322

    2026

  4. [14]

    Single-tree Delineation by Instance Segmentation Using Drone-based Lidar and Multispectral Imagery: a Comparative Study in Various Forest Structures,

    P. Krzystek, et al, "Single-tree Delineation by Instance Segmentation Using Drone-based Lidar and Multispectral Imagery: a Comparative Study in Various Forest Structures," PFG– Journal of Photogrammetry, Remote Sensing and Geoinformation Science, pp. 1-25, 2026. https://doi.or...

    2026 doi

  5. [15]

    Advancing traffic object detection in complex environments: a deep learning object detection approach with vehicle-mounted UAV data for traffic scene perception,

    Y. Liu, et al, "Advancing traffic object detection in complex environments: a deep learning object detection approach with vehicle-mounted UAV data for traffic scene perception," Knowledge and Information Systems, vol. 68, no. 1, 2026. https://doi.org/10.1007/s10115-025- 02621-0

    2026 doi

  6. [16]

    Enhancing Wildfire Preparedness and Response: A Drone Network- Based Early Warning System for Bushfires,

    M. S. A. Reshan, et al, "Enhancing Wildfire Preparedness and Response: A Drone Network- Based Early Warning System for Bushfires," Concurrency and Computation: Practice and Experience, vol. 38, no. 1, 2026. https://doi.org/10.1002/cpe.70528

    2026 doi

  7. [17]

    Rich feature hierarchies for accurate object detection and semantic segmentation,

    R. Girshick, et al, "Rich feature hierarchies for accurate object detection and semantic segmentation," In Proceedings of the IEEE conference on computer vision and pattern recognition, pp.580-587, 2014. https://doi.org/10.1109/CVPR.2014.81

    2014 doi

  8. [18]

    Fast r-cnn,

    R. Girshick, "Fast r-cnn," In Proceedings of the IEEE international conference on computer vision, pp. 1440-1448, 2015. https://doi.org/10.1109/ICCV.2015.169

    2015 doi

  9. [19]

    Faster R-CNN: Towards real-time object detection with region proposal networks,

    S. Ren, et al, "Faster R-CNN: Towards real-time object detection with region proposal networks," Advances in neural information processing systems, 2015. https://doi.org/10.1109/TPAMI.2016.2577031

    2015

  10. [20]

    Mask r-cnn,

    K. He, et al, “Mask r-cnn,” In Proceedings of the IEEE international conference on computer vision, pp. 2961-2969, 2017. https://doi.org/10.1109/ICCV.2017.322

    2017 doi

  11. [21]

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

    J. Redmon, et al, "You only look once: Unified real-time object detection," In Proceedings of the IEEE conference on computer vision and pattern recognition, pp. 779-788, 2016. https://doi.org/10.1109/CVPR.2016.91

    2016 doi

  12. [22]

    YOLO9000: better, faster, stronger,

    J. Redmon, et al, "YOLO9000: better, faster, stronger," In Proceedings of the IEEE conference on computer vision and pattern recognition, pp. 7263-7271, 2017. https://doi.org/10.1109/CVPR.2017.690

    2017 doi

  13. [23]

    Yolov3: An incremental improvement,

    J. Redmon, "Yolov3: An incremental improvement," 2018. arXiv, https://doi.org/10.48550/arXiv.1804.02767

    2018 doi arXiv

  14. [24]

    Yolov4: Optimal speed and accuracy of object detection,

    A. Bochkovskiy, et al, "Yolov4: Optimal speed and accuracy of object detection," arXiv, 2020. https://doi.org/10.48550/arXiv.2004.10934

    2020 doi arXiv

  15. [25]

    Available online: https://github.com/ultralytics/yolov5, (accessed on: 23 November 2022)

    Glenn-jocher. Available online: https://github.com/ultralytics/yolov5, (accessed on: 23 November 2022)

    2022

  16. [26]

    YOLOv6: A single-state object detection framework for industrial applications,

    C. Li, et al, "YOLOv6: A single-state object detection framework for industrial applications," arXiv, 2022. https://doi.org/10.48550/arXiv.2209.02976

    2022 doi arXiv

  17. [27]

    YOLOv7: Trainable bag-of-freebies sets new state-of-the-art for real-time object detectors,

    C. Y. Wang, et al, "YOLOv7: Trainable bag-of-freebies sets new state-of-the-art for real-time object detectors," In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pp. 7464-7475, 2023. https://doi.org/10.1109/CVPR52729.2023.00721

    2023

  18. [28]

    Available online: https://docs.ultralytics.com/models/yolov8 (accessed on: 11 November 2023)

    Glenn-jocher. Available online: https://docs.ultralytics.com/models/yolov8 (accessed on: 11 November 2023)

    2023

  19. [29]

    YOLOv9: Learning what you want to learn using programmable gradient information. European Conference on Computer Vision,

    C. Y. Wang, et al, "YOLOv9: Learning what you want to learn using programmable gradient information. European Conference on Computer Vision," In European conference on computer vision, pp. 1–21, 2024. https://doi.org/10.1007/978-3-031-72751-1_1

    2024 doi

  20. [30]

    YOLOv10: Real-Time End-to-End Object Detection,

    A. Wang, et al, "YOLOv10: Real-Time End-to-End Object Detection," Advances in Neural Information Processing Systems, 2024. https://doi.org/10.52202/079017-3429

    2024 doi

  21. [31]

    YOLOv11: An Overview of the Key Architectural Enhancements,

    R. Khanam, et al, "YOLOv11: An Overview of the Key Architectural Enhancements," arXiv,

  22. [32]

    Ssd: Single shot multibox detector,

    W. Liu, et al, "Ssd: Single shot multibox detector," In European conference on computer vision, pp. 21-37, 2016. https://doi.org/10.1007/978-3-319-46448-0_2

    2016 doi

  23. [33]

    YOLOv12: Attention-Centric Real-Time Object Detectors,

    Y. Tian, et al, "YOLOv12: Attention-Centric Real-Time Object Detectors," arXiv, 2025. https://doi.org/10.48550/arXiv.2502.12524

    2025 doi arXiv

  24. [34]

    Available online: https://github.com/ultralytics/ultralytics, (accessed on: 14 January 2026)

    Glenn-jocher. Available online: https://github.com/ultralytics/ultralytics, (accessed on: 14 January 2026)

    2026

  25. [35]

    LogicAD: Explainable Anomaly Detection via VLM-based Text Feature Extraction,

    E. Jin, et al, "LogicAD: Explainable Anomaly Detection via VLM-based Text Feature Extraction," In Proceedings of the AAAI Conference on Artificial Intelligence, vol. 39, no. 4, pp. 4129-4137, 2025. https://doi.org/10.1609/aaai.v39i4.32433

    2025 doi

  26. [36]

    Focal loss for dense object detection,

    T. Y. Lin, et al, "Focal loss for dense object detection," In Proceedings of the IEEE international conference on computer vision, pp. 2980-2988, 2017. https://doi.org/10.1109/ICCV.2017.324

    2017 doi

  27. [37]

    Single-shot refinement neural network for object detection,

    S. Zhang, et al, "Single-shot refinement neural network for object detection," In Proceedings of the IEEE conference on computer vision and pattern recognition, pp. 4203-4212, 2018. https://doi.org/10.1109/CVPR.2018.00442

    2018

  28. [38]

    Strong and Weak Prompt Engineering for Remote Sensing Image-Text Cross- Modal Retrieval,

    T. Sun, et al, "Strong and Weak Prompt Engineering for Remote Sensing Image-Text Cross- Modal Retrieval," IEEE Journal of Selected Topics in Applied Earth Observations and Remote Sensing, 2025. https://doi.org/10.1109/JSTARS.2025.3534474

    2025

  29. [39]

    GADNet: Improving image–text matching via graph-based aggregation and disentanglement,

    X. Pu, et al, "GADNet: Improving image–text matching via graph-based aggregation and disentanglement," Pattern Recognition, 2025. https://doi.org/10.1016/j.patcog.2024.110900

    2025

  30. [40]

    A System of Multimodal Image-Text Retrieval Based on Pre-Trained Models Fusion,

    Q. Li, et al, "A System of Multimodal Image-Text Retrieval Based on Pre-Trained Models Fusion," Concurrency and Computation: Practice and Experience, vol. 37, no. 3, 2025. https://doi.org/10.1002/cpe.8345

    2025 doi

  31. [41]

    Make It Count: Text-to-Image Generation with an Accurate Number of Objects,

    L. Binyamin, et al, "Make It Count: Text-to-Image Generation with an Accurate Number of Objects," In Proceedings of the Computer Vision and Pattern Recognition Conference, pp. 13242- 13251, 2025. https://doi.org/10.1109/CVPR52734.2025.01236

    2025

  32. [42]

    Local-enhanced representation for text-based person search,

    G. Zhang, et al, "Local-enhanced representation for text-based person search," Pattern Recognition, 2024. https://doi.org/10.1016/j.patcog.2024.111247

    2024

  33. [44]

    An efficient YOLOv12-based framework for detecting extremely small-scale objects,

    A. Chandrashekhar, et al, "An efficient YOLOv12-based framework for detecting extremely small-scale objects," Scientific Reports, 2025. https://doi.org/10.1038/s41598-025-31803-7

    2025 doi

  34. [45]

    Perception-Guided Jailbreak Against Text-to-Image Models,

    Y. Huang, et al, "Perception-Guided Jailbreak Against Text-to-Image Models," In Proceedings of the AAAI Conference on Artificial Intelligence, vol. 39, no. 25, pp. 26238-26247,

  35. [46]

    A text-guided vision model for enhanced recognition of small instances,

    H. K. Jung, "A text-guided vision model for enhanced recognition of small instances," Applied Computer Science, vol. 22, no. 1, pp. 35-46, 2026. https://doi.org/10.35784/acs_7850

    2026 doi

  36. [47]

    YOLO-World: Real-Time Open-Vocabulary Object Detection,

    T. Cheng, et al, "YOLO-World: Real-Time Open-Vocabulary Object Detection," In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pp. 16901- 16911, 2024. https://doi.org/ 10.1109/CVPR52733.2024.01599

    2024

  37. [49]

    CSPNet: A new backbone that can enhance learning capability of CNN,

    C. Y. Wang, et al, "CSPNet: A new backbone that can enhance learning capability of CNN," In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition workshops, pp. 390-391, 2020. https://doi.org/10.1109/CVPRW50498.2020.00203

    2020

  38. [2024]

    https://doi.org/10.48550/arXiv.2410.17725

    doi arXiv

  39. [2025]

    https://doi.org/10.1609/aaai.v39i25.34821

    doi

Pith tools

Reviewed August 1, 2026 · model on record in the stance chip above.