Pith. sign in

REVIEW 1 major objections 1 minor 35 references

Deformable Attention Mechanisms Applied to Object Detection, case of Remote Sensing

T0 review · 1 major / 1 minor · reviewed 2026-08-07 · deepseek-v4-flash

Pith's one-line read The paper reports that Deformable-DETR, with multi-scale deformable attention, achieves the best F1 score and shortest training time among seven detectors on both optical and SAR remote sensing datasets.

desk verdict A usable but under-specified benchmark claiming Deformable-DETR wins on two remote sensing datasets; the core claim needs protocol details and corrected numbers before it can be trusted. read the letter →

arxiv 2505.24489 v1 pith:VV3TZJ6U submitted 2025-05-30 cs.CV cs.AI

classification cs.CVcs.AI
keywords Deformable-DETRdeformableattentionobjectdetectionremotesensingopticalaircraftSARshiptransformerdetectors10-foldcross-validation
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 sets out to show that Deformable-DETR—a transformer detector whose attention mechanism samples a sparse set of points at multiple feature scales instead of attending to every location—is the best all-round choice for object detection on remote sensing images. Using one optical dataset (Pleiades aircraft) and one SAR dataset (ship detection), it compares seven detectors under a common protocol and reports that Deformable-DETR achieves the highest F1 score on both datasets (95.12% and 94.54%) while also requiring the least training time over 12 epochs. It additionally reports that Deformable-DETR beats three remote-sensing-specific transformer detectors on the original SAR ship dataset after 200 epochs of fine-tuning. If these results hold, deformable attention would be a strong, efficient default for detecting small and multi-scale objects in overhead imagery, in both passive optical and active SAR modes.

What carries the argument

The load-bearing object is the deformable attention module introduced by the Deformable-DETR architecture. Instead of computing attention over the full feature map, each query samples $N$ points at each of $M$ feature-map scales, with learnable sampling offsets and attention weights; the output is a weighted sum over these sampled points, written in the paper as $\sum_{m=1}^{M}\sum_{n=1}^{N} A_{mn} W_{mn} f(x_{mn})$. This mechanism simultaneously reduces the computational cost of transformer attention and lets the model adapt to objects of different sizes, which is why the paper attributes both the accuracy gains and the training-time savings to it.

What would settle it

Run the published benchmark notebooks on SSDD and Pleiades Aircraft using exactly the same folds, 12 epochs, COCO-pretrained weights, and augmentation pipeline, and check two numbers: whether Deformable-DETR's average F1 across the ten folds reproduces 94.54% and 95.12%, and whether its training time is still the shortest when all models are measured with identical I/O and logging. A single model whose per-fold metrics are not reproduced, or whose training time was measured without the same data-loading overhead, would falsify the speed-and-accuracy claim as stated.

Watch

Extended reading notes

Core claim

On the paper's own terms, the discovery is that deformable attention, not a larger model or longer training schedule, is what accounts for the best accuracy-speed trade-off. Deformable-DETR's multi-scale deformable attention lets each query concentrate on the most informative points across feature-map levels, which the paper identifies as the mechanism behind its strong performance on small-scale objects and its faster convergence. With COCO-pretrained weights and 12 epochs, the model posts F1 95.12% and mean average precision over IoU thresholds 0.5 to 0.95 of 76.75% on Pleiades Aircraft, and F1 94.54% and 76.14% on SSDD, with the shortest training times in both settings. After 200 epochs of fine-tuning on the original SSDD, Deformable-DETR reaches 98.13% mAP at IoU 0.5 and F1 96.36%, surpassing the specialized remote sensing detectors against which it was compared.

Load-bearing premise

The central claim stands on the assumption that every model was trained and validated under the same protocol, with the reported 10-fold stratified cross-validation actually run as described; if the 80/10/10 split and the 10-fold validation were not properly reconciled, or if models used different augmentation, pretraining, or hyperparameters, the ranking could change.

Editorial extensions

If this is right

  • If Deformable-DETR is genuinely the fastest and most accurate among the seven models at 12 epochs, the same compute budget can support more training epochs or more experiments, so the reported advantage compounds in practice.
  • The result suggests that one deformable-attention architecture can serve both optical and SAR remote sensing, reducing the need for sensor-specific detectors.
  • The strong result on small-scale aircraft and ship instances implies that deformable attention is a useful mechanism for small-object detection in high-resolution overhead imagery.
  • The comparison against remote-sensing-specific transformers indicates that a general-purpose detector with deformable attention can match or beat specialized designs on SAR ship detection.
  • The release of the benchmark code and notebooks would let other teams reproduce the ranking and extend it to new datasets.

Reading between the lines

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

  • A caveat the paper does not state: its headline ranking is a 12-epoch result, so the conclusion that deformable attention trains fastest may not survive longer schedules where other detectors catch up; testing at matching convergence would separate 'faster to converge' from 'better at convergence'.
  • Because the dataset annotations were manually corrected, especially the truncated-object classes, the reported F1 scores partly reflect annotation curation; a sensitivity check on the original versus corrected annotations would show how much of the advantage is due to the attention mechanism itself.
  • A natural testable extension is to feed the same deformable attention into a segmentation head, a direction the paper itself proposes as future work; the mechanism's scale-adaptivity should transfer beyond bounding boxes and can be checked directly with the released code.
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

1 major / 1 minor

Summary. The paper evaluates Deformable-DETR for object detection on two remote sensing datasets, one optical (Pleiades Aircraft) and one SAR (SSDD), and compares it with six other generic detectors (RetinaNet, Faster R-CNN, YOLOv11, DETR, DN-DETR, Conditional DETR, DAB-DETR) and three remote-sensing-specific transformer-based detectors (ASAFE, DAPN, CRTTransSar). The authors report that Deformable-DETR achieves the best F1 scores (95.12% on Pleiades, 94.54% on SSDD) and the shortest reported training time over 12 epochs, and conclude that deformable attention mechanisms are well suited to remote sensing object detection.

Significance. If the benchmark were properly controlled and reproducible, the paper would provide useful evidence that a generic deformable-attention detector can compete with or outperform both CNN-based detectors and remote-sensing-specific architectures while training quickly. The authors also release code and data, which is commendable for reproducibility. However, as presented, the central claim is weakened by protocol ambiguity, an internal inconsistency in the reported training times, and a mismatch between the main comparison and the comparison against specialized remote sensing detectors. The core idea is plausible and the work is potentially valuable, but the evidence does not yet support the strength of the stated conclusions.

major comments (1)
  1. [Table 2, Pleiades Aircraft row for DETR] The reported F1 score for DETR is inconsistent with the reported precision and recall. For precision 93.21 and recall 90.35, the harmonic mean is 2×93.21×90.35/(93.21+90.35) ≈ 91.76, not 91.19 as listed. This suggests an arithmetic or data-handling error in the metrics. Please verify all entries in Table 2 and re-check the F1 calculations for every model, since a single error in the table undermines confidence in the entire numerical comparison.
minor comments (1)
  1. [References] Several references have formatting issues, such as missing spaces after initials (e.g., 'V . V`eque') and missing page ranges; a careful proofread of the reference list would improve the manuscript.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the paper is an empirical benchmark and its conclusions do not reduce to fitted inputs or self-citations.

full rationale

The paper's central claim — that Deformable-DETR achieved the best F1 scores and the shortest training time among the compared detectors — is an empirical benchmark result, not a derivation. No quantity is defined in terms of another quantity it is later said to predict; no fitted parameter is renamed as a prediction; and no uniqueness or ansatz is imported from the authors' prior work. The authors' self-citations (references [1], [2], and [4]) appear in the introduction and related-work survey and are not load-bearing for the experimental conclusion. The evaluation-protocol ambiguities noted in Section 5 (an 80/10/10 split alongside a stratified 10-fold validation, and the discrepancy between the training-time values in the text and Table 2) and the protocol mismatch in Table 3 are legitimate concerns about fairness and reproducibility, but they are correctness risks rather than circular reasoning. Since the result is an external benchmark with no derivation chain that reduces to its own inputs, the circularity score is 0.

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

The central claim rests on the chosen datasets, the training protocol, and the annotation corrections. No invented entities or new physical/mathematical postulates are introduced.

free parameters (3)
  • Learning rate, batch size, optimizer, and other training hyperparameters
    Not reported in the paper; these choices affect all compared models and the reported training times.
  • Number of training epochs (12) for the main comparison = 12
    A single epoch count is applied to all models without evidence that each has converged; this choice may bias the comparison.
  • Manual annotation corrections (truncated aircraft/ship bounding boxes)
    The paper states that bounding boxes were manually corrected, but does not describe the criteria or provide the corrected annotation files.
assumptions (3)
  • domain assumption Pretrained COCO weights are a suitable initialization for remote sensing optical and SAR images
    Used in Section 5 for all models; if remote sensing images are out-of-distribution for COCO, the fine-tuning comparison may be unfair.
  • domain assumption The stratified 10-fold cross-validation is correctly implemented
    The paper describes the protocol ambiguously; the validity of the reported averages depends on a correct implementation.
  • standard math The standard definitions of precision, recall, F1, and mAP are used
    Equations (2)-(6) match standard definitions, though the F1 for DETR in Table 2 is arithmetically inconsistent.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Deformable Attention Mechanisms Applied to Object Detection, case of Remote Sensing." pith.science (2026). https://pith.science/paper/VV3TZJ6U

@misc{pith2026250524489,
  author       = {Pith},
  title        = {Pith review of: Deformable Attention Mechanisms Applied to Object Detection, case of Remote Sensing},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/VV3TZJ6U}},
  note         = {Machine review of arXiv:2505.24489}
}
read the original abstract

Object detection has recently seen an interesting trend in terms of the most innovative research work, this task being of particular importance in the field of remote sensing, given the consistency of these images in terms of geographical coverage and the objects present. Furthermore, Deep Learning (DL) models, in particular those based on Transformers, are especially relevant for visual computing tasks in general, and target detection in particular. Thus, the present work proposes an application of Deformable-DETR model, a specific architecture using deformable attention mechanisms, on remote sensing images in two different modes, especially optical and Synthetic Aperture Radar (SAR). To achieve this objective, two datasets are used, one optical, which is Pleiades Aircraft dataset, and the other SAR, in particular SAR Ship Detection Dataset (SSDD). The results of a 10-fold stratified validation showed that the proposed model performed particularly well, obtaining an F1 score of 95.12% for the optical dataset and 94.54% for SSDD, while comparing these results with several models detections, especially those based on CNNs and transformers, as well as those specifically designed to detect different object classes in remote sensing images.

Figures

Figures reproduced from arXiv: 2505.24489 by the authors.

Figure 1
Figure 1. Deformable-DETR architecture ([18]). 3.1.2 RetinaNet Developed by Facebook AI Research (FAIR) in 2017 [20], RetinaNet is a CNN-based detector. First, a ResNet-50 [21] backbone is deployed for feature extraction, then, a Feature Pyramid Network (FPN) is utilized to create a pyramid of features based on different object scales, the goal behind introducing FPN is to be able to detect objects at different scales. Next, … view at source ↗
Figure 2
Figure 2. Sample images of the two datasets (top: Pleiades Aircraft dataset, bottom: SSDD). [PITH_FULL_IMAGE:figures/full_fig_p004_2.png] view at source ↗
Figure 3
Figure 3. Methodology workflow. part of data augmentations, transformations are performed on the images/annotations of each dataset during the training and validation. The aim behind these transformations, which are horizontal flip, grayscale, i.e. converting a color image from the RGB mode to grayscale, and Gaussian Blur, which involves the application of image blurring based on a Gaussian distribution, is to improve the rob… view at source ↗
Figures from the paper (2 more)
Figure 4
Figure 4. Figure 4: Combined Normalized Loss (Regression + Classification) across models. [PITH_FULL_IMAGE:figures/full_fig_p007_4.png]
Figure 5
Figure 5. Figure 5: Detection results on the two datasets. 6 Conclusion and future work This paper evaluates the performance of deformable attention mechanisms, through Deformable-DETR model, while comparing it with the best-performing models for the task of object detection, whether base…

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

35 extracted references · 27 canonical work pages

  1. [1]

    Energy efficient routing for wireless mesh networks with directional antennas: When q-learning meets ant systems,

    I. Lahsen-Cherif, L. Zitoune, and V . V`eque, “Energy efficient routing for wireless mesh networks with directional antennas: When q-learning meets ant systems,” Ad Hoc Networks, vol. 121, p. 102589, 2021

  2. [2]

    Real-time drone anti-collision avoidance systems: an edge artificial intel- ligence application,

    I. Lahsen-Cherif, H. Liu, and C. Lamy-Bergot, “Real-time drone anti-collision avoidance systems: an edge artificial intel- ligence application,” in 2022 IEEE radar conference (RadarConf22) , pp. 1–6, IEEE, 2022

  3. [3]

    Attention is all you need,

    A. Vaswani, N. Shazeer, N. Parmar, J. Uszkoreit, L. Jones, A. N. Gomez, Ł. Kaiser, and I. Polosukhin, “Attention is all you need,” Advances in neural information processing systems , vol. 30, 2017

  4. [4]

    When machine learning meets geospatial data: A comprehensive geoai review.,

    A. Boutayeb, I. Lahsen-Cherif, and A. El Khadimi, “When machine learning meets geospatial data: A comprehensive geoai review.,”IEEE Journal of Selected Topics in Applied Earth Observations and Remote Sensing , 2025. 8

  5. [5]

    A weakly supervised deep learning framework for sorghum head detection and counting,

    S. Ghosal, B. Zheng, S. C. Chapman, A. B. Potgieter, D. R. Jordan, X. Wang, A. K. Singh, A. Singh, M. Hirafuji, S. Ni- nomiya, et al., “A weakly supervised deep learning framework for sorghum head detection and counting,”Plant Phenomics, 2019

  6. [6]

    Military object detection using multiple information extracted from hyperspectral imagery,

    C. Ke, “Military object detection using multiple information extracted from hyperspectral imagery,” in 2017 International Conference on Progress in Informatics and Computing (PIC) , pp. 124–128, IEEE, 2017

  7. [7]

    Automatic detection of solar photovoltaic arrays in high resolution aerial imagery,

    J. M. Malof, K. Bradbury, L. M. Collins, and R. G. Newell, “Automatic detection of solar photovoltaic arrays in high resolution aerial imagery,” Applied energy, vol. 183, pp. 229–240, 2016

  8. [8]

    A comparative study of deep learning-based models for object detection in remote sensing imagery,

    A. V . Coulson, W. H. Thomas, and C. Wang, “A comparative study of deep learning-based models for object detection in remote sensing imagery,” The International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences, vol. 48, pp. 201–205, 2025

Show all 35 references
  1. [9]

    Ssd: Single shot multibox detector,

    W. Liu, D. Anguelov, D. Erhan, C. Szegedy, S. Reed, C.-Y . Fu, and A. C. Berg, “Ssd: Single shot multibox detector,” in Computer Vision–ECCV 2016: 14th European Conference, Amsterdam, The Netherlands, October 11–14, 2016, Proceed- ings, Part I 14, pp. 21–37, Springer, 2016

  2. [10]

    Transformers in remote sensing: A survey,

    A. A. Aleissaee, A. Kumar, R. M. Anwer, S. Khan, H. Cholakkal, G.-S. Xia, and F. S. Khan, “Transformers in remote sensing: A survey,” Remote Sensing, vol. 15, no. 7, p. 1860, 2023

  3. [11]

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

    A. Dosovitskiy, L. Beyer, A. Kolesnikov, D. Weissenborn, X. Zhai, T. Unterthiner, M. Dehghani, M. Minderer, G. Heigold, S. Gelly, et al. , “An image is worth 16x16 words: Transformers for image recognition at scale,” arXiv preprint arXiv:2010.11929, 2020

  4. [12]

    Ao2-detr: Arbitrary-oriented object detection transformer,

    L. Dai, H. Liu, H. Tang, Z. Wu, and P. Song, “Ao2-detr: Arbitrary-oriented object detection transformer,” IEEE Transac- tions on Circuits and Systems for Video Technology, vol. 33, no. 5, pp. 2342–2356, 2022

  5. [13]

    Oriented object detection with transformer,

    T. Ma, M. Mao, H. Zheng, P. Gao, X. Wang, S. Han, E. Ding, B. Zhang, and D. Doermann, “Oriented object detection with transformer,” arXiv preprint arXiv:2106.03146, 2021

  6. [14]

    Road extraction method of remote sensing image based on deformable attention transformer,

    L. Zhao, J. Zhang, X. Meng, W. Zhou, Z. Zhang, and C. Peng, “Road extraction method of remote sensing image based on deformable attention transformer,” Symmetry, vol. 16, no. 4, p. 468, 2024

  7. [15]

    Mabdt: Multi-scale attention boosted deformable transformer for remote sensing image dehazing,

    J. Ning, J. Yin, F. Deng, and L. Xie, “Mabdt: Multi-scale attention boosted deformable transformer for remote sensing image dehazing,” Signal Processing, vol. 229, p. 109768, 2025

  8. [16]

    Pr-deformable detr: Detr for remote sensing object detection,

    Y . Chen, B. Liu, and L. Yuan, “Pr-deformable detr: Detr for remote sensing object detection,”IEEE Geoscience and Remote Sensing Letters, 2024

  9. [17]

    Refined deformable-detr for sar target detection and radio signal detection,

    Z. Li and X. Zhou, “Refined deformable-detr for sar target detection and radio signal detection,” Remote Sensing, vol. 17, no. 8, p. 1406, 2025

  10. [18]

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

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

  11. [19]

    End-to-end object detection with transform- ers,

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

  12. [20]

    Focal loss for dense object detection,

    T.-Y . Lin, P. Goyal, R. Girshick, K. He, and P. Doll´ar, “Focal loss for dense object detection,” in Proceedings of the IEEE international conference on computer vision , pp. 2980–2988, 2017

  13. [21]

    Deep residual learning for image recognition,

    K. He, X. Zhang, S. Ren, and J. Sun, “Deep residual learning for image recognition,” inProceedings of the IEEE conference on computer vision and pattern recognition , pp. 770–778, 2016

  14. [22]

    Faster r-cnn: Towards real-time object detection with region proposal networks,

    S. Ren, K. He, R. Girshick, and J. Sun, “Faster r-cnn: Towards real-time object detection with region proposal networks,” Advances in neural information processing systems , vol. 28, 2015

  15. [23]

    Yolov11: An overview of the key architectural enhancements,

    R. Khanam and M. Hussain, “Yolov11: An overview of the key architectural enhancements,” arXiv preprint arXiv:2410.17725, 2024

  16. [24]

    Dn-detr: Accelerate detr training by introducing query denoising,

    F. Li, H. Zhang, S. Liu, J. Guo, L. M. Ni, and L. Zhang, “Dn-detr: Accelerate detr training by introducing query denoising,” in Proceedings of the IEEE/CVF conference on computer vision and pattern recognition , pp. 13619–13627, 2022

  17. [25]

    Conditional detr for fast training convergence,

    D. Meng, X. Chen, Z. Fan, G. Zeng, H. Li, Y . Yuan, L. Sun, and J. Wang, “Conditional detr for fast training convergence,” in Proceedings of the IEEE/CVF international conference on computer vision , pp. 3651–3660, 2021. 9

  18. [26]

    Dab-detr: Dynamic anchor boxes are better queries for detr,

    S. Liu, F. Li, H. Zhang, X. Yang, X. Qi, H. Su, J. Zhu, and L. Zhang, “Dab-detr: Dynamic anchor boxes are better queries for detr,” arXiv preprint arXiv:2201.12329, 2022

  19. [27]

    Airbus aircrafts detection sample dataset

    Airbus, “Airbus aircrafts detection sample dataset.” https://www.kaggle.com/datasets/airbusgeo/ airbus-aircrafts-sample-dataset , 2021. Retrieved from Kaggle

  20. [28]

    Sar ship detection dataset (ssdd): Official release and comprehensive data analysis,

    T. Zhang, X. Zhang, J. Li, X. Xu, B. Wang, X. Zhan, Y . Xu, X. Ke, T. Zeng, H. Su,et al., “Sar ship detection dataset (ssdd): Official release and comprehensive data analysis,” Remote Sensing, vol. 13, no. 18, p. 3690, 2021

  21. [29]

    Microsoft coco: Common objects in context,

    T.-Y . Lin, M. Maire, S. Belongie, J. Hays, P. Perona, D. Ramanan, P. Doll´ar, and C. L. Zitnick, “Microsoft coco: Common objects in context,” in Computer vision–ECCV 2014: 13th European conference, zurich, Switzerland, September 6-12, 2014, proceedings, part v 13 , pp. 740–75...

  22. [30]

    An adaptive sample assignment strategy based on feature enhancement for ship detection in sar images,

    H. Shi, Z. Fang, Y . Wang, and L. Chen, “An adaptive sample assignment strategy based on feature enhancement for ship detection in sar images,” Remote Sensing, vol. 14, no. 9, p. 2238, 2022

  23. [31]

    Dense attention pyramid networks for multi-scale ship detection in sar images,

    Z. Cui, Q. Li, Z. Cao, and N. Liu, “Dense attention pyramid networks for multi-scale ship detection in sar images,” IEEE Transactions on Geoscience and Remote Sensing , vol. 57, no. 11, pp. 8983–8997, 2019

  24. [32]

    A local-sparse-information-aggregation transformer with explicit contour guidance for sar ship detection,

    H. Shi, B. Chai, Y . Wang, and L. Chen, “A local-sparse-information-aggregation transformer with explicit contour guidance for sar ship detection,” Remote Sensing, vol. 14, no. 20, p. 5247, 2022

  25. [33]

    Imagenet: A large-scale hierarchical image database,

    J. Deng, W. Dong, R. Socher, L.-J. Li, K. Li, and L. Fei-Fei, “Imagenet: A large-scale hierarchical image database,” in 2009 IEEE conference on computer vision and pattern recognition , pp. 248–255, Ieee, 2009

  26. [34]

    Training data-efficient image transformers & distillation through attention,

    H. Touvron, M. Cord, M. Douze, F. Massa, A. Sablayrolles, and H. J ´egou, “Training data-efficient image transformers & distillation through attention,” in International conference on machine learning , pp. 10347–10357, PMLR, 2021

  27. [35]

    Swin transformer: Hierarchical vision transformer using shifted windows,

    Z. Liu, Y . Lin, Y . Cao, H. Hu, Y . Wei, Z. Zhang, S. Lin, and B. Guo, “Swin transformer: Hierarchical vision transformer using shifted windows,” in Proceedings of the IEEE/CVF international conference on computer vision , pp. 10012–10022, 2021. 10

Pith tools

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