Pith. sign in

REVIEW 3 major objections 5 minor 45 references

DenoDet V2: Phase-Amplitude Cross Denoising for SAR Object Detection

T0 review · 3 major / 5 minor · reviewed 2026-08-05 · deepseek-v4-flash

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

Pith's one-line read DenoDet V2 claims that cleaning SAR features in the frequency domain, by letting phase and amplitude spectra swap attention tokens, lifts detection accuracy to 56.71% mAP on SARDet-100K, 1.7% over its baseline and 0.8% over its predecessor,

desk verdict Solid incremental SAR detection paper with a novel frequency-domain attention module, but the central phase-robustness claim is untested at the feature level and the reported gains could be noise. read the letter →

arxiv 2508.09392 v1 pith:B6MU2YQV submitted 2025-08-12 cs.CV

Kang Ni , Minrui Zou , Yuxuan Li , Xiang Li , Kehua Guo , Ming-Ming Cheng , Yimian Dai This is my paper
classification cs.CV
keywords SARobjectdetectionfrequency-domaindenoisingphase-amplitudecrossmodulationself-attentionspecklenoisediscreteFouriertransformtokenexchange
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 claims that coherent speckle noise in SAR imagery can be suppressed inside a detector by denoising features in the frequency domain rather than the spatial domain. It introduces DenoDet V2, which transforms feature maps with a 2D DFT, splits them into amplitude and phase spectra, and lets the two spectra guide each other's denoising through a band-wise self-attention module called Phase-Amplitude Token Exchange. The central assertion is that phase information is more noise-robust, so it can act as a reference to clean the amplitude spectrum, and the cleaned amplitude then improves the phase. If correct, the result is a plug-and-play module that raises detection accuracy on three SAR benchmarks while cutting model complexity roughly in half compared with DenoDet V1.

What carries the argument

The DFTDeno module: a plug-in block that applies a 2D discrete Fourier transform to deep feature maps, computes amplitude and phase spectra, modulates them with a band-wise partition self-attention, recombines them, and returns to space via inverse DFT. Its core is Phase-Amplitude Token Exchange (PATE): within each frequency band, the attention tokens of the amplitude and phase spectra are swapped, so phase guides amplitude denoising and amplitude guides phase refinement. The band-wise partition (stride 8) keeps attention local in frequency while reducing parameters.

What would settle it

Train DenoDet V2 on SARDet-100K with the PATE token exchange disabled but with attention heads of equal parameter count; if mAP stays at 56.7%, the claimed mechanism is not the source of the gain. Alternatively, inject controlled speckle into the phase versus the amplitude of feature maps; if phase corruption degrades mAP as much as amplitude corruption, the phase-robustness premise fails.

Watch

Extended reading notes

Core claim

On the largest SAR detection benchmark, SARDet-100K, DenoDet V2 reports 56.71% mAP, 1.7% above its GFL baseline and 0.8% above DenoDet V1, with about half the parameters of V1 and negligible added FLOPs; it also reports state-of-the-art results on SAR-Aircraft-1.0 (69.93%) and AIR-SARShip-1.0 (73.98%). The mechanism behind the gains is a reciprocal denoising loop: DFT-transformed feature maps are decomposed into amplitude and phase, partitioned into frequency bands, and the self-attention queries, keys, and values are exchanged between phase and amplitude within each band. The paper argues that phase, being less affected by speckle, serves as a stable reference that soft-thresholds noise in

Load-bearing premise

The load-bearing premise is that phase spectra of DFT-transformed deep feature maps remain noise-robust inside the network, just as image-space phase is; if speckle corrupts feature-level phase too, the cross-denosing loop has no advantage over ordinary attention.

Editorial extensions

If this is right

  • If the claim holds, frequency-domain feature denoising can be added to generic SAR detectors as a plug-in without redesigning the detector's backbone or head.
  • The largest gains appear on small and medium objects (AP_S 51.45, AP_M 68.75), which are most vulnerable to speckle, suggesting the mechanism specifically helps low-signature targets.
  • Achieving state-of-the-art results on three diverse SAR benchmarks indicates the reciprocal phase-amplitude exchange transfers across object classes and imaging conditions.
  • Halving the parameter count relative to DenoDet V1 while improving accuracy implies the cross-modulation itself, not extra capacity, drives the reported gain.
  • The ablation sweep over partition strides shows global frequency attention (stride 1) hurts performance, so local band-wise attention is a necessary ingredient of the design.

Reading between the lines

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

  • Editorial extension: the same reciprocal phase-amplitude exchange could generalize to other coherent-imaging modalities where speckle or multiplicative noise dominates, such as ultrasound or sonar, by porting DFTDeno into detectors trained on those modalities.
  • Editorial extension: the phase-robustness premise is directly testable in-feature: corrupt the phase spectrum of DFT-transformed feature maps versus the amplitude spectrum with synthetic speckle; if the claim is right, phase corruption should degrade mAP less.
  • Editorial extension: the paper does not explore iterating the exchange more than once; a multi-pass phase-amplitude refinement loop could either compound the denoising benefit or reveal diminishing returns.
  • Editorial extension: the same attention-as-cross-denoising idea may extend to other transform domains, such as wavelets or DCT, where amplitude-like and phase-like components also play asymmetric roles under noise.
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

3 major / 5 minor

Summary. DenoDet V2 proposes a plug-in DFTDeno module for SAR object detection. After applying a 2D DFT to deep feature maps, the amplitude and phase spectra are decomposed and modulated by band-wise self-attention (BPSA), with a Phase-Amplitude Token Exchange (PATE) mechanism in which query/key/value tokens are exchanged between amplitude and phase groups within corresponding frequency bands. The module is inserted into a GFL detector and evaluated on SARDet-100K, SAR-Aircraft-1.0, and AIR-SARShip-1.0. The paper reports state-of-the-art results, notably 56.71% mAP on SARDet-100K, +1.7% over GFL and +0.8% over DenoDet V1 while roughly halving the parameter count. Ablations show incremental gains from orthogonal phase decomposition, phase/amplitude refinement, band partitioning, and token exchange.

Significance. If the reported results hold, DenoDet V2 is a meaningful empirical advance: it ranks first on the largest public SAR detection benchmark, improves over the authors' own V1 at half the parameters, and is evaluated on three datasets with code release. The ablation structure is logical and the efficiency claim is well supported. However, the paper's distinctive scientific claim—that phase robustness transfers to DFT-transformed deep features and that cross-spectral token exchange is the cause of the gains—is not directly tested. The headline differences are small (0.3 mAP for token exchange, 0.8 mAP vs. V1), and the lack of error bars and test-set hyperparameter selection make the SOTA claim fragile as currently presented.

major comments (3)
  1. [Section 3.3 / Fig. 1] The central premise that phase is robust to noise at the feature level is untested. The only direct evidence for phase robustness is in natural-image space (Fig. 1 and Chen et al. 2021a), not for DFT-transformed deep features inside the SAR detector. Table 4 compares attention applied to phase versus amplitude, but this changes the processing module and its parameters, not the phase/amplitude content itself. The +0.3 mAP from token exchange in Table 6 is the only piece of evidence for the cross-denosing mechanism, and it could be due to extra attention interactions. To support the mechanistic claim, please add feature-domain perturbation experiments: randomize or swap feature-level phase while preserving amplitude (and vice versa), or hold the attention topology fixed and vary only which spectral component supplies Q/K/V.
  2. [Section 4.3, Table 5] The band partition stride is selected on the test set. The paper reports that stride=8 gives the best SARDet-100K mAP and then adopts stride=8 for the final SOTA comparison on the same benchmark. This is test-set fitting: the reported 56.71% mAP is not an unbiased estimate of generalization performance. Please select hyperparameters using the validation fold (the dataset has an 8:1:1 split) or use nested/held-out evaluation, and report the final test result only after fixing the stride.
  3. [Section 4.3, Tables 4 and 6; Table 1] The key empirical claims lack uncertainty quantification. The headline comparisons are small: +0.3 mAP for token exchange over no-exchange, +0.8 mAP over DenoDet V1, and +1.7 mAP over GFL. Table 4's phase-versus-amplitude difference is only 0.6 mAP. All numbers appear to be single runs with no error bars, multiple seeds, or significance testing. On COCO-style mAP, differences of this size can easily arise from optimization noise. Please report means and standard deviations over at least three seeds, or bootstrap confidence intervals, for the main comparisons and ablations.
minor comments (5)
  1. [Title page / Abstract] Typo: 'Augest' should be 'August'.
  2. [Section 4.1] The resized image dimensions appear as garbled characters ('��� � ���') in the provided text; please ensure all image sizes, batch sizes, and FLOPs are rendered correctly.
  3. [Table 2 caption] The caption says 'SAR-Aircraft' while the text refers to 'SAR-Aircraft-1.0'; please make the dataset naming consistent.
  4. [Section 3.2, Eqs. (12)-(14)] The symbol d is used both for the embedding/channel dimension and for the number of band groups; please disambiguate (e.g., use d_model and G).
  5. [Supplementary Tables S7-S9] Several entries in the supplementary tables are blank or missing (e.g., AP_L and per-class values for many methods). Please complete the tables or mark unavailable values explicitly.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the SOTA claims are benchmarked on external public datasets and the phase-amplitude mechanism is evaluated by ablations, not derived from its own outputs.

full rationale

The paper's central performance claims are measured on independent public benchmarks (SARDet-100K, SAR-Aircraft-1.0, AIR-SARShip-1.0) against 25 external baselines, so the reported mAP numbers are not constructed from the model's own fitted values or from a self-referential definition. The phase-robustness premise is imported from an external non-author paper (Chen et al. 2021a) and then tested at feature level via ablations in Table 4 that compare phase-only, amplitude-only, and joint attention modulation; the conclusion that phase refinement helps more is an empirical comparison, not an identity or a fitted parameter renamed as a prediction. The PATE module is a fixed architectural choice whose contribution is measured by controlled ablations (Tabs. 4-6); no parameter is fitted to a subset of data and then reported as a prediction of the same data. The authors' prior DenoDet V1 and the SARDet-100K benchmark appear as comparison/baseline elements but are not used to derive the reported gains. The skeptic's concern that feature-level phase perturbation was not directly performed is a validity or confound concern, not a circularity: nothing in the paper's equations makes the claimed conclusion equivalent to its inputs by construction. Therefore no circular step can be exhibited under the paper's own derivations.

Assumptions & free parameters 1 free parameters · 4 assumptions · 2 invented entities

The central claims rest on a few domain assumptions (phase robustness, separability of DFT components, attention as denoising) and one free hyperparameter (partition stride) selected on the test benchmark. There are no invented physical entities; the listed entities are neural network modules. The support for these modules is internal to the paper.

free parameters (1)
  • band partition stride (h, w) = 8
    The stride that partitions the DFT spectrum into groups was chosen by comparing values {1, 2, 4, 8, 16} on the SARDet-100K test benchmark and selecting the maximum mAP (Table 5). This is model selection on the evaluation set, so the final number is partially fit to the reported result.
assumptions (4)
  • domain assumption Phase information is more robust to noise than amplitude information and preserves structural integrity.
    Invoked in the Introduction and Section 3.3 as the foundation for phase-guided denoising. Cited from Chen et al. 2021a for natural images, but assumed to hold for SAR feature maps.
  • domain assumption DFT on feature maps yields meaningful amplitude and phase components that can be manipulated independently.
    The method applies 2D DFT to intermediate CNN features and treats amplitude and phase as separable signals (Eqs. 5-8).
  • domain assumption Attention can act as a denoiser in the frequency domain.
    Carried over from DenoDet V1 (Dai et al. 2024) and the 'attention as feature denoising' concept.
  • ad hoc to paper Band-wise partitioning preserves frequency locality that is useful for denoising.
    The paper argues frequencies are centrally symmetric (Fig. S10) and groups them into bands, but the optimal stride is an empirical choice (Table 5).
invented entities (2)
  • Phase-Amplitude Token Exchange (PATE) module
    purpose: Allows phase and amplitude spectra to mutually modulate each other in shared frequency bands via swapped Q/K/V tokens.
    The paper introduces this module as the core novelty, but its effectiveness is demonstrated only by internal ablations (Tables 4-6) on the same benchmarks used for the final claims; there is no external falsifiable prediction.
  • Band-wise Partition Self-Attention (BPSA)
    purpose: Decomposes global frequency-domain attention into parallelizable sub-band computations to preserve frequency distributions and reduce dimensionality.
    This is an architectural component whose benefits are shown only through the paper's own ablations (Table 5); no independent test is provided.

how reviews work

0 comments
Cite this review

Pith. "Pith review of DenoDet V2: Phase-Amplitude Cross Denoising for SAR Object Detection." pith.science (2026). https://pith.science/paper/B6MU2YQV

@misc{pith2026250809392,
  author       = {Pith},
  title        = {Pith review of: DenoDet V2: Phase-Amplitude Cross Denoising for SAR Object Detection},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/B6MU2YQV}},
  note         = {Machine review of arXiv:2508.09392}
}
read the original abstract

One of the primary challenges in Synthetic Aperture Radar (SAR) object detection lies in the pervasive influence of coherent noise. As a common practice, most existing methods, whether handcrafted approaches or deep learning-based methods, employ the analysis or enhancement of object spatial-domain characteristics to achieve implicit denoising. In this paper, we propose DenoDet V2, which explores a completely novel and different perspective to deconstruct and modulate the features in the transform domain via a carefully designed attention architecture. Compared to DenoDet V1, DenoDet V2 is a major advancement that exploits the complementary nature of amplitude and phase information through a band-wise mutual modulation mechanism, which enables a reciprocal enhancement between phase and amplitude spectra. Extensive experiments on various SAR datasets demonstrate the state-of-the-art performance of DenoDet V2. Notably, DenoDet V2 achieves a significant 0.8\% improvement on SARDet-100K dataset compared to DenoDet V1, while reducing the model complexity by half. The code is available at https://github.com/GrokCV/GrokSAR.

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

45 extracted references · 38 canonical work pages

  1. [1]

    write newline

    " write newline "" before.all 'output.state := FUNCTION n.dashify 't := "" t empty not t #1 #1 substring "-" = t #1 #2 substring "--" = not "--" * t #2 global.max substring 't := t #1 #1 substring "-" = "-" * t #2 global.max substring 't := while if t #1 #1 substring * t #2 global.max substring 't := if while FUNCTION format.date year duplicate empty "emp...

  2. [2]

    Cascade R-CNN : High quality object detection and instance segmentation

    Zhaowei Cai and Nuno Vasconcelos. Cascade R-CNN : High quality object detection and instance segmentation. IEEE Transactions on Pattern Analysis and Machine Intelligence, 43 0 (5): 0 1483--1498, 2021

    work page 2021

  3. [3]

    End-to-end object detection with transformers

    Nicolas Carion, Francisco Massa, Gabriel Synnaeve, Nicolas Usunier, Alexander Kirillov, and Sergey Zagoruyko. End-to-end object detection with transformers. In European conference on computer vision, pages 213--229. Springer, 2020

    work page 2020

  4. [4]

    Amplitude-phase recombination: Rethinking robustness of convolutional neural networks in frequency domain

    Guangyao Chen, Peixi Peng, Li Ma, Jia Li, Lin Du, and Yonghong Tian. Amplitude-phase recombination: Rethinking robustness of convolutional neural networks in frequency domain. In 2021 IEEE/CVF International Conference on Computer Vision (ICCV), pages 448--457, 2021 a

    work page 2021

  5. [5]

    You only look one-level feature

    Qiang Chen, Yingming Wang, Tong Yang, Xiangyu Zhang, Jian Cheng, and Jian Sun. You only look one-level feature. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 13039--13048, 2021 b

    work page 2021

  6. [6]

    Disentangle your dense object detector

    Zehui Chen, Chenhongyi Yang, Qiaofei Li, Feng Zhao, Zheng-Jun Zha, and Feng Wu. Disentangle your dense object detector. In ACM International Conference on Multimedia, pages 4939--4948, 2021 c

    work page 2021

  7. [7]

    Denodet: Attention as deformable multi-subspace feature denoising for target detection in sar images

    Yimian Dai, Minrui Zou, Yuxuan Li, Xiang Li, Kang Ni, and Jian Yang. Denodet: Attention as deformable multi-subspace feature denoising for target detection in sar images. IEEE Transactions on Aerospace and Electronic Systems (TAES), 2024

    work page 2024

  8. [8]

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

    work page 2021

Show all 45 references
  1. [9]

    YOLOX : Exceeding yolo series in 2021

    Zheng Ge, Songtao Liu, Feng Wang, Zeming Li, and Jian Sun. YOLOX : Exceeding yolo series in 2021. arXiv preprint, arXiv:2107.08430, 2021

    2021 arXiv

  2. [10]

    X. Ke, X. Zhang, T. Zhang, J. Shi, and S. Wei. SAR ship detection based on an improved faster R-CNN using deformable convolution. In Proceedings of the IEEE International Geoscience and Remote Sensing Symposium (IGARSS), pages 3565--3568, 2021

    2021

  3. [11]

    Probabilistic anchor assignment with IoU prediction for object detection

    Kang Kim and Hee Seok Lee. Probabilistic anchor assignment with IoU prediction for object detection. In Computer Vision--ECCV 2020: 16th European Conference, Glasgow, UK, August 23--28, 2020, Proceedings, Part XXV 16, pages 355--371. Springer, 2020

    2020

  4. [12]

    A novel multidimensional domain deep learning network for SAR ship detection

    Dong Li, Quanhuan Liang, Hongqing Liu, Qinghua Liu, Haijun Liu, and Guisheng Liao. A novel multidimensional domain deep learning network for SAR ship detection. IEEE Transactions on Geoscience and Remote Sensing , 60: 0 1--13, 2022 a

    2022

  5. [13]

    D. Li, Q. Liang, H. Liu, Q. Liu, H. Liu, and G. Liao. A novel multidimensional domain deep learning network for SAR ship detection. IEEE Transactions on Geoscience and Remote Sensing, 60: 0 5203213, 2022 b

    2022

  6. [14]

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

    Xiang Li, Wenhai Wang, Lijun Wu, Shuo Chen, Xiaolin Hu, Jun Li, Jinhui Tang, and Jian Yang. Generalized focal loss: Learning qualified and distributed bounding boxes for dense object detection. In Advances in Neural Information Processing Systems, 2020

    2020

  7. [15]

    Large selective kernel network for remote sensing object detection

    Yuxuan Li, Qibin Hou, Zhaohui Zheng, Ming-Ming Cheng, Jian Yang, and Xiang Li. Large selective kernel network for remote sensing object detection. In IEEE International Conference on Computer Vision, pages 16794--16805, 2023

    2023

  8. [16]

    SARDet -100K: towards open-source benchmark and toolKit for large-scale SAR object detection

    Yuxuan Li, Xiang Li, Weijie Li, Qibin Hou, Li Liu, Ming-Ming Cheng, and Jian Yang. SARDet -100K: towards open-source benchmark and toolKit for large-scale SAR object detection . In Advances in Neural Information Processing Systems, pages 128430--128461, Brookline, MA, USA, 202...

    2024

  9. [17]

    Sm3det: A unified model for multi-modal remote sensing object detection

    Yuxuan Li, Xiang Li, Yunheng Li, Zhang Yicheng, Yimian Dai, Qibin Hou, Ming-Ming Cheng, and Jian Yang. Sm3det: A unified model for multi-modal remote sensing object detection. arXiv, 2024 b

    2024

  10. [18]

    Focal loss for dense object detection

    Tsung-Yi Lin, Priya Goyal, Ross Girshick, Kaiming He, and Piotr Dollar. Focal loss for dense object detection. In Proceedings of the IEEE International Conference on Computer Vision (ICCV), pages 2980--2988, 2017

    2017

  11. [19]

    DAB-DETR : Dynamic anchor boxes are better queries for DETR

    Shilong Liu, Feng Li, Hao Zhang, Xiao Yang, Xianbiao Qi, Hang Su, Jun Zhu, and Lei Zhang. DAB-DETR : Dynamic anchor boxes are better queries for DETR . arXiv preprint arXiv:2201.12329, 2022 a

    2022 arXiv

  12. [20]

    DAB-DETR : Dynamic anchor boxes are better queries for DETR

    Shilong Liu, Feng Li, Hao Zhang, Xiao Yang, Xianbiao Qi, Hang Su, Jun Zhu, and Lei Zhang. DAB-DETR : Dynamic anchor boxes are better queries for DETR . In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 3651--3660, 2022 b

    2022

  13. [21]

    A ConvNet for the 2020s

    Zhuang Liu, Hanzi Mao, Chao-Yuan Wu, Christoph Feichtenhofer, Trevor Darrell, and Saining Xie. A ConvNet for the 2020s. In IEEE Conference on Computer Vision and Pattern Recognition, pages 11976--11986, 2022 c

    2022

  14. [22]

    Grid R-CNN

    Xin Lu, Buyu Li, Yuxin Yue, Quanquan Li, and Junjie Yan. Grid R-CNN . In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 7363--7372, 2019

    2019

  15. [23]

    Libra R-CNN : Towards balanced learning for object detection

    Jiangmiao Pang, Kai Chen, Jianping Shi, Huajun Feng, Wanli Ouyang, and Dahua Lin. Libra R-CNN : Towards balanced learning for object detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 821--830, 2019

    2019

  16. [24]

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

    Shaoqing Ren, Kaiming He, Ross Girshick, and Jian Sun. Faster R-CNN : Towards real-time object detection with region proposal networks. Advances in neural information processing systems, 28, 2015

    2015

  17. [25]

    H. Shen, C. Zhou, J. Li, and Q. Yuan. SAR image despeckling employing a recursive deep CNN prior. IEEE Transactions on Geoscience and Remote Sensing, 59 0 (1): 0 273--286, 2021

    2021

  18. [26]

    Y. Sun, X. Sun, Z. Wang, and K. Fu. Oriented ship detection based on strong scattering points network in large-scale SAR images. IEEE Transactions on Geoscience and Remote Sensing, 60: 0 5218018, 2021

    2021

  19. [27]

    FCOS : Fully convolutional one-stage object detection

    Zhi Tian, Chunhua Shen, Hao Chen, and Tong He. FCOS : Fully convolutional one-stage object detection. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 9627--9636, 2019

    2019

  20. [28]

    Attention is all you need

    Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, ukasz Kaiser, and Illia Polosukhin. Attention is all you need. Advances in neural information processing systems, 30, 2017

    2017

  21. [29]

    Automatic sar ship detection based on multifeature fusion network in spatial and frequency domains

    Shiyu Wang, Zhanchuan Cai, and Jieyu Yuan. Automatic sar ship detection based on multifeature fusion network in spatial and frequency domains. IEEE Transactions on Geoscience and Remote Sensing, 61: 0 1--11, 2023

    2023

  22. [30]

    Pyramid vision transformer: A versatile backbone for dense prediction without convolutions

    Wenhai Wang, Enze Xie, Xiang Li, Deng-Ping Fan, Kaitao Song, Ding Liang, Tong Lu, Ping Luo, and Ling Shao. Pyramid vision transformer: A versatile backbone for dense prediction without convolutions. In Proceedings of the IEEE/CVF international conference on computer vision, pa...

    2021

  23. [31]

    ConvNeXt V2 : Co-designing and scaling convnets with masked autoencoders

    Sanghyun Woo, Shoubhik Debnath, Ronghang Hu, Xinlei Chen, Zhuang Liu, In So Kweon, and Saining Xie. ConvNeXt V2 : Co-designing and scaling convnets with masked autoencoders. In IEEE Conference on Computer Vision and Pattern Recognition, pages 16133--16142, 2023

    2023

  24. [32]

    AIR-SARShip-1.0 : High-resolution SAR ship detection dataset

    SUN Xian, WANG Zhirui, SUN Yuanrui, DIAO Wenhui, ZHANG Yue, and FU Kun. AIR-SARShip-1.0 : High-resolution SAR ship detection dataset. Journal of Radars, 8 0 (6): 0 852--863, 2019

    2019

  25. [33]

    K. Xu, M. Qin, F. Sun, Y. Wang, Y.-K. Chen, and F. Ren. Learning in the frequency domain. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 1737--1746, 2020

    2020

  26. [34]

    RepPoints : Point set representation for object detection

    Ze Yang, Shaohui Liu, Han Hu, Liwei Wang, and Stephen Lin. RepPoints : Point set representation for object detection. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 9657--9666, 2019

    2019

  27. [35]

    Dynamic R-CNN : Towards high quality object detection via dynamic training

    Hongkai Zhang, Hong Chang, Bingpeng Ma, Naiyan Wang, and Xilin Chen. Dynamic R-CNN : Towards high quality object detection via dynamic training. In Computer Vision--ECCV 2020: 16th European Conference, Glasgow, UK, August 23--28, 2020, Proceedings, Part XV 16, pages 260--275. ...

    2020

  28. [36]

    VarifocalNet : An IoU -aware dense object detector

    Haoyang Zhang, Ying Wang, Feras Dayoub, and Niko Sunderhauf. VarifocalNet : An IoU -aware dense object detector. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 8514--8523, 2021

    2021

  29. [37]

    Shifeng Zhang, Cheng Chi, Yongqiang Yao, Zhen Lei, and Stan Z. Li. Bridging the gap between anchor-based and anchor-free detection via adaptive training sample selection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 9759--9...

    2020

  30. [38]

    Rsar: Restricted state angle resolver and rotated sar benchmark

    Xin Zhang, Xue Yang, Yuxuan Li, Jian Yang, Ming-Ming Cheng, and Xiang Li. Rsar: Restricted state angle resolver and rotated sar benchmark. In IEEE Conference on Computer Vision and Pattern Recognition, pages 7416--7426, 2025

    2025

  31. [39]

    Sar ship detection based on end-to-end morphological feature pyramid network

    Congxia Zhao, Xiongjun Fu, Jian Dong, Rui Qin, Jiayun Chang, and Ping Lang. Sar ship detection based on end-to-end morphological feature pyramid network. IEEE Journal of Selected Topics in Applied Earth Observations and Remote Sensing, 15: 0 4599--4611, 2022

    2022

  32. [40]

    J. Zhao, Z. Zhang, W. Yu, and T.-K. Truong. A cascade coupled convolutional neural network guided visual attention method for ship detection from SAR images. IEEE Access, 6: 0 50693--50708, 2018

    2018

  33. [41]

    SAR-AIRcraft-1.0 : High-resolution SAR aircraft detection and recognition dataset

    WANG Zhirui, KANG Yuzhuo, ZENG Xuan, WANG Yuelei, ZHANG Ting, and SUN Xian. SAR-AIRcraft-1.0 : High-resolution SAR aircraft detection and recognition dataset. Journal of Radars, 12 0 (4): 0 906--922, 2023

    2023

  34. [42]

    Objects as points

    Xingyi Zhou, Dequan Wang, and Philipp Krähenbühl. Objects as points. arXiv preprint arXiv:1904.07850, 2019

    1904 arXiv

  35. [43]

    PVT-SAR : An arbitrarily oriented SAR ship detector with pyramid vision transformer

    Yue Zhou, Xue Jiang, Guozheng Xu, Xue Yang, Xingzhao Liu, and Zhou Li. PVT-SAR : An arbitrarily oriented SAR ship detector with pyramid vision transformer. IEEE Journal of Selected Topics in Applied Earth Observations and Remote Sensing, 16: 0 291--305, 2022

    2022

  36. [44]

    Autoassign: Differentiable label assignment for dense object detection

    Benjin Zhu, Jianfeng Wang, Zhengkai Jiang, Fuhang Zong, Songtao Liu, Zeming Li, and Jian Sun. Autoassign: Differentiable label assignment for dense object detection. arXiv preprint arXiv:2007.03496, 2020

    2007 arXiv

  37. [45]

    Deformable DETR : Deformable transformers for end-to-end object detection

    Xizhou Zhu, Weijie Su, Lewei Lu, Bin Li, Xiaogang Wang, and Jifeng Dai. Deformable DETR : Deformable transformers for end-to-end object detection. In International Conference on Learning Representations, 2021

    2021

Pith tools

Reviewed August 5, 2026 · model on record in the stance chip above.