Pith. sign in

REVIEW 5 major objections 7 minor 48 references

Language-guided Learning for Object Detection Tackling Multiple Variations in Aerial Images

T0 review · 5 major / 7 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 LANGO uses language-guided learning to handle both scene-level and instance-level variations in aerial object detection.

desk verdict A solid, incremental aerial-detection paper whose language-semantics story is plausible but under-tested; the reported gains deserve a referee, not a desk reject. read the letter →

arxiv 2505.23193 v1 pith:V674MWZA submitted 2025-05-29 cs.CV

Sungjune Park , Hyunjun Kim , Beomchan Park , Yong Man Ro This is my paper · ORCID
classification cs.CV
keywords aerialobjectdetectionlanguage-guidedlearningvisualsemanticreasonerrelationlossscene-levelvariationinstance-leveldroneimagerytransformer
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 proposes LANGO, an aerial-image object detector that uses language as a guide to handle scene-level variation (weather, illumination) and instance-level variation (viewpoint, scale) at the same time. A visual semantic reasoner reads global scene context from an image and sharpens the encoder's features to match it, while a relation learning loss trains visual object features to reproduce the similarity structure of sentence embeddings for object descriptions. On UAVDT the framework reports 57.9 AP with a ResNet-50 backbone and 58.6 AP with ResNet-101, and on VisDrone it reports 64.4 and 64.8 AP, above the earlier methods it compares against. Ablation results attribute the improvement to both components.

What carries the argument

The load-bearing mechanism is language-guided learning with two complementary parts. The visual semantic reasoner is a vision-to-language cross-attention block: image features serve as queries against a scene context prompt, the output is added back into the encoder features, and a frozen large language model generates a scene description during training to supervise the reasoner. The relation learning loss computes, for each object category, a softmax-normalized similarity vector from the category exemplar's sentence embedding to all other instance embeddings, including variants like 'small' or 'side view'; it then trains the visual object feature of that category to produce the same distribution over those embeddings, minimized by KL divergence. This transfers the invariance of the sentence embedding space to the visual feature space.

What would settle it

Train the full LANGO pipeline twice on UAVDT, once with the real sentence embeddings and once with the same embeddings randomly permuted per category; if the AP gap between the two runs nearly disappears, the relation loss is not exploiting language semantics. A second check is to evaluate on a test subset with matched viewpoint and scale and see whether the relation loss still gives its full ablation gain when instance-level variation is minimal.

Watch

Extended reading notes

Core claim

LANGO's central claim is that language representations can donate their invariance to a visual detector. Sentences such as 'a side view photo of a small vehicle in a foggy day' are encoded by a pretrained sentence model, and these embeddings stay clustered by object category even when the described viewpoint, scale, or weather changes. The detector is trained so each visual object feature's cosine-similarity distribution over the full set of language instances matches the distribution of its category exemplar, using KL divergence. In parallel, a vision-to-language cross-attention module combined with a frozen large language model (used only in training) nudges the encoder features to encode scene conditions. The authors' conclusion is that the resulting detector recognizes objects accurately when weather, altitude, or camera angle change.

Load-bearing premise

The method assumes that sentence embeddings of object descriptions stay clustered by category under viewpoint, scale, and weather changes, and that forcing visual features to match that similarity structure improves detection rather than distorting it.

Editorial extensions

If this is right

  • On UAVDT, LANGO reports 57.9 AP with ResNet-50 and 58.6 AP with ResNet-101, exceeding the published comparison methods listed in the paper.
  • On VisDrone, the same components report 64.4 and 64.8 AP, ahead of the earlier methods the paper compares against.
  • The two design components are additive: the reasoner alone improves the baseline, the relation loss alone improves it, and together they give the best result on both datasets.
  • The language-guided components transfer to another transformer detector, DDQ-DETR, raising its UAVDT AP from 48.2 to 56.0 with ResNet-50 and from 48.6 to 57.2 with Swin-L.
  • Inference stays near real time: 0.057 seconds per image on an RTX 3090, only 0.007 seconds more than the base detector.

Reading between the lines

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

  • A direct test of the mechanism would replace the sentence embedding model with random vectors; if AP does not drop, the relation loss's gain is not actually coming from language robustness.
  • The scene reasoner's benefit could be partly regularization rather than semantic understanding; pairing random scene descriptions with images in the reasoner loss would separate those effects.
  • Since VisDrone day/night labels are pseudo-labels from another model, the reasoner's gain on VisDrone may depend on the quality of those labels; using manual annotations would test this.
  • The same language-relation objective could be applied to oriented bounding-box detection, where viewpoint variation is continuous and current detectors struggle with angle changes.
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 / 7 minor

Summary. The paper proposes LANGO, a language-guided object detection framework for aerial images. It consists of two components: a visual semantic reasoner that uses a vision-to-language cross-attention module and a frozen LLM during training to interpret scene-level conditions (weather, view, altitude), and a relation learning loss that aligns visual object features to the similarity structure of language representations of object categories and their viewpoint/scale variants. The framework is built on RT-DETR and evaluated on UAVDT and VisDrone, where it reports state-of-the-art AP (57.9/58.6 on UAVDT and 64.4/64.8 on VisDrone with R50/R101). Ablations on both datasets show that each component improves performance over the baseline.

Significance. If the empirical results are reproducible, the paper makes a useful contribution: it demonstrates a practical way to jointly address scene- and instance-level variations in aerial detection, with consistent gains across two datasets and two backbones, and the inference overhead over its RT-DETR base is small (0.007 s/img). The idea of using the robustness of sentence embeddings as a training target for visual features is interesting and potentially transferable. However, the paper does not release code, leaves key training details unspecified, and does not isolate the semantic contribution of the relation loss, so the significance of the mechanism itself is not yet established.

major comments (5)
  1. [Section III.C, Eq. (3)] The total training objective in Eq. (3) includes only L_cls, L_bbox, and L_R. However, Section III.A describes feeding the reasoner's output features to a frozen LLM together with a scene context prompt and guiding the LLM to generate a scene context description, which implies an additional language-modeling loss on the reasoner. Please specify the exact loss used to train the reasoner and add it to the objective, or explain how the reasoner is trained through the detection losses only.
  2. [Section III.B, Eq. (2)] The relation loss in Eq. (2) is averaged over NC categories, yet the paper also introduces extra non-target categories (e.g., ground, building, shadow) that have no ground-truth annotations and therefore no visual object features o_i. It is unclear how the loss is computed for these categories. If the loss is applied only to the target categories, the notation and averaging need to be revised; if it is applied to all NC, the mechanism for obtaining visual features for non-target categories should be described.
  3. [Section III.B, Fig. 4/5] The central explanatory claim is that language representations are robust to viewpoint and scale variations and that this robustness transfers to visual features through L_R. The ablations show that L_R helps, but they do not distinguish between 'language semantics help' and 'generic category-separation regularization helps.' Please add a control experiment that replaces the language embeddings with random or permuted category anchors of the same dimension, while keeping the loss identical; this would isolate the semantic contribution and support the stated mechanism.
  4. [Section IV.A.2] The VisDrone scene context descriptions are generated from pseudo day/night labels obtained with Meteor, but the accuracy of these pseudo labels is not reported or validated. Since the visual semantic reasoner is trained to predict these labels, label noise could affect the observed gains. Please report the pseudo-label accuracy on a sample or compare with a small manually annotated subset, and discuss the potential impact of noise.
  5. [Section III.B] The preparation of the language representation set Z via contrastive training of the learnable categorical prompts is underspecified: the contrastive loss formulation, the number of training iterations, the learning rate, the number of variant descriptions N per category, and the exact attribute combinations are not provided. Without these details, the method is not fully reproducible.
minor comments (7)
  1. [Eq. (2)] 'Kullback-Leiber' should be 'Kullback-Leibler'.
  2. [Throughout] The dataset name is inconsistently written as 'UA VDT' and 'UAVDT'; please standardize.
  3. [Fig. 5 caption] 'street lampshadow' should be 'street lamp shadow'.
  4. [Section IV.B] 'Mistral-7B-Intsruct-v0.3' is a typo for 'Mistral-7B-Instruct-v0.3'.
  5. [Section III.A] '5 ×5 convolution layers' should be '5×5 convolution layer' or clarify whether multiple layers are used.
  6. [Tables I-III] The paper does not report the number of independent runs or error bars for the key results; please state whether results are from a single run and, if feasible, report variance over multiple runs.
  7. [Table V] Table V would be easier to interpret if the same-backbone baseline comparisons were separated from comparisons with methods using different backbones, since the latter are not directly comparable.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: LANGO's detection gains are measured on external benchmarks and the language-guided targets are fixed before detector training.

full rationale

The central claim is LANGO's detection AP on UAVDT and VisDrone, which is evaluated with standard protocols on held-out test and validation splits. Neither the relation learning loss (Eq. 2) nor the visual semantic reasoner is fitted to those test labels: the language similarity targets R_i^l are computed from pretrained sentence embeddings and contrastively trained categorical prompts before detector training, with no feedback from detection predictions. The visual semantic reasoner is supervised by dataset annotations on UAVDT and by Meteor pseudo labels on VisDrone, which are external sources independent of the detection loss. Thus no reported AP reduces by construction to a fitted input or to the language-derived targets. The paper's assertion that language representations are robust to viewpoint and scale is supported only by t-SNE and lacks a random-embedding control, but that is an interpretability and experimental-validity concern rather than circularity. The one self-citation (Meteor, [29]) is used as an auxiliary pseudo-labeling tool for day/night conditions on VisDrone and is not load-bearing for the state-of-the-art claim; it provides external evidence from a separately trained model and does not raise the circularity score.

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

The method relies on standard deep learning components and on empirical assumptions about language representations and pseudo labels. The learned categorical prompts and hand-selected instance descriptions are the main fitted ingredients.

free parameters (4)
  • Learnable categorical prompts c_i = Not reported
    Trained via unspecified contrastive learning on language representations (Eq. 1, Section III.B); final values are not reported, so the relation distribution targets depend on unstated optimization.
  • Relation loss weight = 1.0 (implicit)
    L_final = L_cls + L_bbox + L_R with no weighting hyperparameter reported (Section III.C); no ablation over the weight is shown.
  • Instance description set per category = N not specified
    The variant descriptions for each category (view, scale, weather combinations) are hand-curated (Section IV.B); the exact list is not provided, so reproduction depends on undisclosed choices.
  • Extra non-target categories = ground, building, tree, traffic sign, street lamp, shadow
    These categories are added to the language representation set (Section V.C); the selection is hand-chosen and its effect on the loss is not ablated.
assumptions (4)
  • domain assumption Pretrained sentence embeddings (all-mpnet-base-v2) capture category relationships robust to viewpoint and scale.
    Core motivation for the relation loss (Section III.B); Fig. 4 shows t-SNE but no quantitative robustness test.
  • domain assumption Mistral-7B can generate accurate scene context descriptions from the visual semantic reasoner's output features.
    The LLM is used as a teacher during training (Section III.A); its outputs guide the reasoner, but no analysis of description quality is provided.
  • domain assumption Meteor pseudo labels for daytime/nighttime in VisDrone are reliable.
    Section IV.A.2 uses pseudo labels from Meteor [29] to supervise the scene reasoner on VisDrone; accuracy is not reported.
  • ad hoc to paper Extra non-target categories help the detector distinguish targets from backgrounds.
    Section V.C adds these categories to the representation set; the choice is not ablated or justified beyond a claim.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Language-guided Learning for Object Detection Tackling Multiple Variations in Aerial Images." pith.science (2026). https://pith.science/paper/V674MWZA

@misc{pith2026250523193,
  author       = {Pith},
  title        = {Pith review of: Language-guided Learning for Object Detection Tackling Multiple Variations in Aerial Images},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/V674MWZA}},
  note         = {Machine review of arXiv:2505.23193}
}
read the original abstract

Despite recent advancements in computer vision research, object detection in aerial images still suffers from several challenges. One primary challenge to be mitigated is the presence of multiple types of variation in aerial images, for example, illumination and viewpoint changes. These variations result in highly diverse image scenes and drastic alterations in object appearance, so that it becomes more complicated to localize objects from the whole image scene and recognize their categories. To address this problem, in this paper, we introduce a novel object detection framework in aerial images, named LANGuage-guided Object detection (LANGO). Upon the proposed language-guided learning, the proposed framework is designed to alleviate the impacts from both scene and instance-level variations. First, we are motivated by the way humans understand the semantics of scenes while perceiving environmental factors in the scenes (e.g., weather). Therefore, we design a visual semantic reasoner that comprehends visual semantics of image scenes by interpreting conditions where the given images were captured. Second, we devise a training objective, named relation learning loss, to deal with instance-level variations, such as viewpoint angle and scale changes. This training objective aims to learn relations in language representations of object categories, with the help of the robust characteristics against such variations. Through extensive experiments, we demonstrate the effectiveness of the proposed method, and our method obtains noticeable detection performance improvements.

Figures

Figures reproduced from arXiv: 2505.23193 by the authors.

Figure 1
Figure 1. The examples illustrate that there exist multiple variations. (a) shows the scene-level variations ( [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. The overall architecture of the proposed method. While the object detection framework takes an input aerial image and extracts image features, we [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗
Figure 3
Figure 3. The details of the visual semantic reasoner consisting of vision-to [PITH_FULL_IMAGE:figures/full_fig_p004_3.png] view at source ↗
Figures from the paper (3 more)
Figure 4
Figure 4. Figure 4: (a) shows t-SNE feature visualization results demonstrating that language instance representations are gathered together by object category, rather than [PITH_FULL_IMAGE:figures/full_fig_p005_4.png]
Figure 5
Figure 5. Figure 5: The t-SNE visualization of language instance representations for object [PITH_FULL_IMAGE:figures/full_fig_p007_5.png]
Figure 6
Figure 6. Figure 6: The qualitative visualization results on severe environmental conditions from UAVDT test set. The upper and lower figures show the detection results [PITH_FULL_IMAGE:figures/full_fig_p008_6.png]

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

48 extracted references · 40 canonical work pages

  1. [1]

    Krmaro: Aerial detection of small-size ground moving objects using kinematic regularization and matrix rank optimization,

    A. ElTantawy and M. S. Shehata, “Krmaro: Aerial detection of small-size ground moving objects using kinematic regularization and matrix rank optimization,” IEEE Transactions on Circuits and Systems for Video Technology, vol. 29, no. 6, pp. 1672–1686, 2018

    work page 2018

  2. [2]

    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,” in Proceedings of the AAAI Conference on Artificial Intelligence, vol. 36, no. 1, 2022, pp. 1026–1033

    work page 2022

  3. [3]

    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,” in Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , 2023, pp. 13 435–13 444

    work page 2023

  4. [4]

    Hierarchical mask prompting and robust integrated regression for oriented object detection,

    Y . Yao, G. Cheng, C. Lang, X. Yuan, X. Xie, and J. Han, “Hierarchical mask prompting and robust integrated regression for oriented object detection,” IEEE Transactions on Circuits and Systems for Video Technology, 2024

    work page 2024

  5. [5]

    Scale optimization using evolutionary reinforcement learning for object detection on drone imagery,

    J. Zhang, X. Yang, W. He, J. Ren, Q. Zhang, Y . Zhao, R. Bai, X. He, and J. Liu, “Scale optimization using evolutionary reinforcement learning for object detection on drone imagery,” in Proceedings of the AAAI Conference on Artificial Intelligence , vol. 38, no. 1, 2024, pp. 410–418

    work page 2024

  6. [6]

    Learning temporary block- based bidirectional incongruity-aware correlation filters for efficient uav object tracking,

    F. Lin, C. Fu, Y . He, F. Guo, and Q. Tang, “Learning temporary block- based bidirectional incongruity-aware correlation filters for efficient uav object tracking,” IEEE Transactions on Circuits and Systems for Video Technology, vol. 31, no. 6, pp. 2160–2174, 2020

    work page 2020

  7. [7]

    Centric probability- based sample selection for oriented object detection,

    Y . Yao, G. Cheng, C. Lang, X. Xie, and J. Han, “Centric probability- based sample selection for oriented object detection,” IEEE Transactions on Circuits and Systems for Video Technology , 2024

    work page 2024

  8. [8]

    Multi-task learning for uav aerial object detection in foggy weather condition,

    W. Fang, G. Zhang, Y . Zheng, and Y . Chen, “Multi-task learning for uav aerial object detection in foggy weather condition,” Remote Sensing, vol. 15, no. 18, p. 4617, 2023

    work page 2023

Show all 48 references
  1. [9]

    Coderainnet: Collaborative deraining network for drone-view object detection in rainy weather conditions,

    Y . Xi, W. Jia, Q. Miao, J. Feng, X. Liu, and F. Li, “Coderainnet: Collaborative deraining network for drone-view object detection in rainy weather conditions,” Remote Sensing, vol. 15, no. 6, p. 1487, 2023

    2023

  2. [10]

    Visual perception in the human brain: How the brain perceives and understands real-world scenes,

    C. G. Bartnik and I. I. Groen, “Visual perception in the human brain: How the brain perceives and understands real-world scenes,” in Oxford Research Encyclopedia of Neuroscience , 2023

    2023

  3. [11]

    Tph-yolov5: Improved yolov5 based on transformer prediction head for object detection on drone- captured scenarios,

    X. Zhu, S. Lyu, X. Wang, and Q. Zhao, “Tph-yolov5: Improved yolov5 based on transformer prediction head for object detection on drone- captured scenarios,” in Proceedings of the IEEE/CVF International Conference on Computer Vision , 2021, pp. 2778–2788

    2021

  4. [12]

    Tiny object detection via regional cross self-attention network,

    K. Cheng, H. Cui, H. A. Ghafoor, H. Wan, Q. Mao, and Y . Zhan, “Tiny object detection via regional cross self-attention network,” IEEE Transactions on Circuits and Systems for Video Technology , 2022

    2022

  5. [13]

    Plug-and-play robust aerial object detection under hazy conditions,

    W. Wu, H. Chang, Z. Chen, and Z. Li, “Plug-and-play robust aerial object detection under hazy conditions,” IEEE Journal of Selected Topics in Applied Earth Observations and Remote Sensing , 2024. 10

    2024

  6. [14]

    Detclip: Dictionary-enriched visual-concept paralleled pre- training for open-world detection,

    L. Yao, J. Han, Y . Wen, X. Liang, D. Xu, W. Zhang, Z. Li, C. Xu, and H. Xu, “Detclip: Dictionary-enriched visual-concept paralleled pre- training for open-world detection,” Advances in Neural Information Processing Systems, vol. 35, pp. 9125–9138, 2022

    2022

  7. [15]

    Text-driven traffic anomaly detection with temporal high-frequency modeling in driving videos,

    R. Liang, Y . Li, J. Zhou, and X. Li, “Text-driven traffic anomaly detection with temporal high-frequency modeling in driving videos,” IEEE Transactions on Circuits and Systems for Video Technology , 2024

    2024

  8. [16]

    Learning domain-aware detection head with prompt tuning,

    H. Li, R. Zhang, H. Yao, X. Song, Y . Hao, Y . Zhao, L. Li, and Y . Chen, “Learning domain-aware detection head with prompt tuning,” Advances in Neural Information Processing Systems , vol. 36, 2024

    2024

  9. [17]

    Dst-det: Open- vocabulary object detection via dynamic self-training,

    S. Xu, X. Li, S. Wu, W. Zhang, Y . Tong, and C. C. Loy, “Dst-det: Open- vocabulary object detection via dynamic self-training,” IEEE Transactions on Circuits and Systems for Video Technology , 2024

    2024

  10. [18]

    Llms meet vlms: Boost open vocabulary object detection with fine-grained descriptors,

    S. Jin, X. Jiang, J. Huang, L. Lu, and S. Lu, “Llms meet vlms: Boost open vocabulary object detection with fine-grained descriptors,” in The Twelfth International Conference on Learning Representations , 2023

    2023

  11. [19]

    Generative region- language pretraining for open-ended object detection,

    L. Chuang, J. Yi, Q. Lizhen, Y . Zehuan, and C. Jianfei, “Generative region- language pretraining for open-ended object detection,” in Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , 2024

    2024

  12. [20]

    Lenna: Language enhanced reasoning detection assistant,

    F. Wei, X. Zhang, A. Zhang, B. Zhang, and X. Chu, “Lenna: Language enhanced reasoning detection assistant,” arXiv preprint arXiv:2312.02433, 2023

    2023 arXiv

  13. [21]

    Detgpt: Detect what you need via reasoning,

    R. Pi, J. Gao, S. Diao, R. Pan, H. Dong, J. Zhang, L. Yao, J. Han, H. Xu, L. Kong et al., “Detgpt: Detect what you need via reasoning,” in The 2023 Conference on Empirical Methods in Natural Language Processing , 2023

    2023

  14. [22]

    Blip-2: Bootstrapping language- image pre-training with frozen image encoders and large language models,

    J. Li, D. Li, S. Savarese, and S. Hoi, “Blip-2: Bootstrapping language- image pre-training with frozen image encoders and large language models,” in International Conference on Machine Learning . PMLR, 2023, pp. 19 730–19 742

    2023

  15. [23]

    Vicuna: An open-source chatbot impressing gpt-4 with 90%* chatgpt quality,

    W.-L. Chiang, Z. Li, Z. Lin, Y . Sheng, Z. Wu, H. Zhang, L. Zheng, S. Zhuang, Y . Zhuang, J. E. Gonzalez et al., “Vicuna: An open-source chatbot impressing gpt-4 with 90%* chatgpt quality,” See https://vicuna. lmsys. org (accessed 14 April 2023) , vol. 2, no. 3, p. 6, 2023

    2023

  16. [24]

    Grounding dino: Marrying dino with grounded pre-training for open-set object detection,

    S. Liu, Z. Zeng, T. Ren, F. Li, H. Zhang, J. Yang, C. Li, J. Yang, H. Su, J. Zhu et al., “Grounding dino: Marrying dino with grounded pre-training for open-set object detection,” arXiv preprint arXiv:2303.05499 , 2023

    2023 arXiv

  17. [25]

    Clip the gap: A single domain generalization approach for object detection,

    V . Vidit, M. Engilberge, and M. Salzmann, “Clip the gap: A single domain generalization approach for object detection,” in Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , 2023, pp. 3219–3229

    2023

  18. [26]

    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,” in Proceedings of the European Conference on Computer Vision, 2018, pp. 370–386

    2018

  19. [27]

    Detection and tracking meet drones challenge,

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

    2021

  20. [28]

    Delving into robust object detection from unmanned aerial vehicles: A deep nuisance disentanglement approach,

    Z. Wu, K. Suresh, P. Narayanan, H. Xu, H. Kwon, and Z. Wang, “Delving into robust object detection from unmanned aerial vehicles: A deep nuisance disentanglement approach,” in Proceedings of the IEEE/CVF International Conference on Computer Vision , 2019, pp. 1201–1210

    2019

  21. [29]

    Meteor: Mamba-based traversal of rationale for large language and vision models,

    B.-K. Lee, C. W. Kim, B. Park, and Y . M. Ro, “Meteor: Mamba-based traversal of rationale for large language and vision models,”arXiv preprint arXiv:2405.15574, 2024

    2024 arXiv

  22. [30]

    Microsoft coco: Common objects in context,

    T.-Y . Lin, M. Maire, S. Belongie, J. Hays, P. Perona, D. Ramanan, P. Dollár, and C. L. Zitnick, “Microsoft coco: Common objects in context,” in Computer Vision–ECCV 2014: 13th European Conference, Zurich, Switzerland, September 6-12, 2014, Proceedings, Part V 13 . Springer, 2...

    2014

  23. [31]

    Pytorch: An imperative style, high-performance deep learning library,

    A. Paszke, S. Gross, F. Massa, A. Lerer, J. Bradbury, G. Chanan, T. Killeen, Z. Lin, N. Gimelshein, L. Antiga et al. , “Pytorch: An imperative style, high-performance deep learning library,” Advances in Neural Information Processing Systems , vol. 32, 2019

    2019

  24. [32]

    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,” 2023

    2023

  25. [33]

    Dense distinct query for end-to-end object detection,

    S. Zhang, X. Wang, J. Wang, J. Pang, C. Lyu, W. Zhang, P. Luo, and K. Chen, “Dense distinct query for end-to-end object detection,” in Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2023, pp. 7329–7338

    2023

  26. [34]

    Mistral 7b,

    A. Q. Jiang, A. Sablayrolles, A. Mensch, C. Bamford, D. S. Chaplot, D. d. l. Casas, F. Bressand, G. Lengyel, G. Lample, L. Saulnier et al., “Mistral 7b,” arXiv preprint arXiv:2310.06825 , 2023

    2023 arXiv

  27. [35]

    Mistral-7b-instruct-v0.3,

    A. Jiang, A. Sablayrolles, A. Tacnet, A. Roux, A. Mensch, A. Herblin- Stoop, B. Bout, B. de Monicault, B. Savary, Bam4d, C. Feldman, D. S. Chaplot, D. de las Casas, E. Arcelin, E. B. Hanna, E. Metzger, G. Lengyel, G. Bour, G. Lample, H. Rajaona, J.-M. Delignon, J. Li, J. Murke...

    2024

  28. [36]

    Mpnet: Masked and permuted pre-training for language understanding,

    K. Song, X. Tan, T. Qin, J. Lu, and T.-Y . Liu, “Mpnet: Masked and permuted pre-training for language understanding,” Advances in Neural Information Processing Systems , vol. 33, pp. 16 857–16 867, 2020, https: //huggingface.co/sentence-transformers/all-mpnet-base-v2 (accessed...

    2020

  29. [37]

    Detecting small objects using a channel-aware deconvolutional network,

    K. Duan, D. Du, D. H. Qi, and Q. Huang, “Detecting small objects using a channel-aware deconvolutional network,” IEEE Transactions on Circuits and Systems for Video Technology, vol. 30, no. 6, pp. 1639–1652, 2019

    2019

  30. [38]

    Guided attention network for object detection and counting on drones,

    C. YuanQiang, D. Du, L. Zhang, L. Wen, W. Wang, Y . Wu, and S. Lyu, “Guided attention network for object detection and counting on drones,” in Proceedings of the 28th ACM International Conference on Multimedia , 2020, pp. 709–717

    2020

  31. [39]

    Improving multiscale object detection with off-centered semantics refinement,

    X. Tang, Q. Yang, D. Xiong, Y . Xie, H. Wang, and R. Li, “Improving multiscale object detection with off-centered semantics refinement,” IEEE Transactions on Circuits and Systems for Video Technology , vol. 32, no. 10, pp. 6888–6899, 2022

    2022

  32. [40]

    Training domain-invariant object detector faster with feature replay and slow learner,

    C. Lee, J. Seo, and H. Jung, “Training domain-invariant object detector faster with feature replay and slow learner,” in Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , 2021, pp. 1172–1181

    2021

  33. [41]

    Spotnet: Self-attention multi-task network for object detection,

    H. Perreault, G.-A. Bilodeau, N. Saunier, and M. Héritier, “Spotnet: Self-attention multi-task network for object detection,” in 2020 17th Conference on Computer and Robot Vision . IEEE, 2020, pp. 230–237

    2020

  34. [42]

    Flsl: Feature-level self-supervised learning,

    Q. Su, A. Netchaev, H. Li, and S. Ji, “Flsl: Feature-level self-supervised learning,” Advances in Neural Information Processing Systems , vol. 36, 2023

    2023

  35. [43]

    Focus-and- detect: A small object detection framework for aerial images,

    O. C. Koyun, R. K. Keser, I. B. Akkaya, and B. U. Töreyin, “Focus-and- detect: A small object detection framework for aerial images,” Signal Processing: Image Communication , vol. 104, p. 116675, 2022

    2022

  36. [44]

    Towards resolving the challenge of long-tail distribution in uav images for object detection,

    W. Yu, T. Yang, and C. Chen, “Towards resolving the challenge of long-tail distribution in uav images for object detection,” in Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision , 2021, pp. 3258–3267

    2021

  37. [45]

    Fldet: Faster and lighter aerial object detector,

    S. Wang, K. Liu, J. Huang, and X. Li, “Fldet: Faster and lighter aerial object detector,” IEEE Transactions on Circuits and Systems for Video Technology, 2024

    2024

  38. [46]

    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

  39. [47]

    Ogmn: Occlusion-guided multi-task network for object detection in uav images,

    X. Li, W. Diao, Y . Mao, P. Gao, X. Mao, X. Li, and X. Sun, “Ogmn: Occlusion-guided multi-task network for object detection in uav images,” ISPRS Journal of Photogrammetry and Remote Sensing , vol. 199, pp. 242–257, 2023

    2023

  40. [48]

    Pareto refocusing for drone-view object detection,

    J. Leng, M. Mo, Y . Zhou, C. Gao, W. Li, and X. Gao, “Pareto refocusing for drone-view object detection,” IEEE Transactions on Circuits and Systems for Video Technology, vol. 33, no. 3, pp. 1320–1334, 2022

    2022

Pith tools

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