Pith. sign in

REVIEW 4 major objections 5 minor 42 references

SAR-NAS: Lightweight SAR Object Detection with Neural Architecture Search

T0 review · 4 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 Automatic channel-width search tunes YOLOv10 for SAR object detection, beating the base model at lower cost.

desk verdict A clean, honest application of one-shot NAS to YOLOv10 for SAR detection; the central claim rests on single-run mAP deltas that could be seed noise, so the result is plausible but not yet established. read the letter →

arxiv 2509.01279 v1 pith:P53HC4G7 submitted 2025-09-01 cs.CV

Xinyi Yu , Zhiwei Lin , Yongtao Wang This is my paper
classification cs.CV
keywords SyntheticApertureRadarobjectdetectionneuralarchitecturesearchone-shotNASchannel-widthYOLOv10SARDet-100Klightweight
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

This paper tries to establish that automatic search over backbone channel widths can make a strong real-time detector better and cheaper on Synthetic Aperture Radar imagery, without SAR-specific hand engineering. It applies an evolutionary one-shot neural architecture search to YOLOv10, exploring about four million channel configurations under parameter and FLOP constraints. On the SARDet-100K benchmark, the searched small and nano models beat their YOLOv10 counterparts in mAP while using fewer FLOPs and parameters. The authors position this as the first use of NAS for SAR object detection and as evidence that architecture search can replace manual tuning for this sensor domain.

What carries the argument

The load-bearing mechanism is a one-shot weight-sharing search over a channel-width search space. A supernet containing all candidate backbones is trained with the sandwich rule, sampling the largest, smallest, and two intermediate sub-networks each training step. During evolutionary search, candidate architectures inherit weights from this supernet, receive a validation mAP, and are refined by mutation and crossover under hardware constraints; the top candidates are then retrained from scratch. The claim rests on these inherited-weight validation scores being reliable stand-ins for retrained performance.

What would settle it

Take the top candidates selected by inherited-weight validation plus a random sample of lower-ranked candidates, retrain each from scratch under identical settings, and compare final SARDet-100K mAP. If the rank correlation between proxy scores and retrained scores is weak or negative, the reported gain is an artifact of the one-shot proxy rather than a property of the searched architecture.

Watch

Extended reading notes

Core claim

The paper's central claim is that a NAS-searched channel configuration for YOLOv10's backbone yields a lightweight SAR detector that is both more accurate and less costly than the original YOLOv10 and than previous SAR detection models. On SARDet-100K, SAR-NAS-S reaches 69.27 mAP with 20.17 GFLOPs and 7.07M parameters, compared with YOLOv10-S at 68.84 mAP, 21.43 GFLOPs, and 7.22M parameters; SAR-NAS-N reaches 60.18 mAP versus YOLOv10-N at 59.47 mAP while also reducing cost. The search space scales each backbone layer's channel width by 0.25, 0.5, 0.75, or 1.0, yielding roughly four million candidates, and the evolutionary search is constrained by total parameter count and FLOPs. The authors

Load-bearing premise

During search, candidate networks are scored using weights borrowed from one jointly trained parent network instead of being trained on their own; the whole search assumes those proxy scores rank candidates the same way full retraining would.

Editorial extensions

If this is right

  • If the central claim holds, general-purpose real-time detectors can be adapted to SAR without hand-designed SAR-specific modules, shortening the design cycle.
  • Channel-width-only NAS is sufficient to recover nontrivial accuracy-efficiency gains, suggesting that search over width is a low-risk first step for other sensor domains.
  • The searched channel patterns—full width near the neck, reduced width elsewhere—can be read as a design heuristic for lightweight SAR backbones, even without running NAS.
  • The reported search cost of under 10 GPU-days makes this style of NAS practical for teams that cannot afford training thousands of isolated architectures.
  • The same one-shot pipeline should transfer to other YOLO-family detectors and to other remote sensing benchmarks, since nothing in the search space is SAR-specific.

Reading between the lines

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

  • The paper leaves implicit that the strongest contribution may be procedural rather than the absolute gain: it shows how to bolt NAS onto a strong baseline with modest compute, and the +0.43 mAP gain may matter more as evidence of feasibility than as a head-to-head win.
  • Because only backbone widths are searched, a natural untested extension is to widen the search space to block type, depth, and neck channels; larger gains might appear once those dimensions are included.
  • The architecture trend analysis suggests a testable rule: keeping full width at stride positions that feed the feature pyramid is worth more than uniform width scaling; this could be validated by hand-building a width pattern from the trend alone.
  • The reliability of the whole approach depends on one-shot proxy rankings, and the paper does not measure that correlation; a simple diagnostic would be to retrain several low-ranked candidates and check whether they actually underperform.
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 introduces SAR-NAS, a one-shot neural architecture search framework that optimizes the channel widths of the YOLOv10 backbone for SAR object detection. The search space contains roughly four million candidate sub-networks; an evolutionary algorithm with hardware-aware constraints selects architectures using validation accuracy under inherited supernet weights, and the top 10 are retrained from scratch. On SARDet-100K, the resulting SAR-NAS-N and SAR-NAS-S models are reported to achieve 60.18% and 69.27% mAP, respectively, with lower FLOPs and parameter counts than the corresponding YOLOv10 baselines. The paper also compares against a range of generic and SAR-specific detectors, and analyzes the channel-ratio patterns of the searched architectures.

Significance. If the central claims hold, the paper would be a useful early demonstration that NAS can improve a strong real-time detector on a large SAR benchmark, with modest but consistent gains in both accuracy and efficiency. The controlled comparison against YOLOv10-N/S is the right experimental design, and the paper explicitly frames the contribution as a feasibility study rather than a new NAS algorithm. The search-space design and the hardware-aware evolutionary search with retraining of top candidates are sensible and reproducible in principle. However, the current evidence for the central claim is weak: the reported gains are small, are derived from single runs with a best-of-ten selection protocol, and rely on an unvalidated weight-sharing proxy. These issues are addressable but currently undermine confidence in the paper's main conclusion.

major comments (4)
  1. [§4.1 and Table 2] The central comparison reports SAR-NAS-N at 60.18% vs YOLOv10-N at 59.47% (+0.71) and SAR-NAS-S at 69.27% vs YOLOv10-S at 68.84% (+0.43). No error bars, standard deviations, or number of seeds are given for any model. §4.1 states that after search, the top 10 architectures are retrained and the best one is selected. This best-of-10 protocol upwardly biases the SAR-NAS result, while the YOLOv10 baselines appear to be single runs. For gains in the 0.4–0.7 mAP range, seed variance alone could explain the difference. Please report mean±std over at least three independent training runs for both SAR-NAS and YOLOv10, and either use the mean of the top-10 retrained models or match the selection protocol for the baseline.
  2. [Algorithm 1 and §3.1] The evolutionary search ranks candidate architectures by Validate(arch, W, Dval), where W are inherited supernet weights. The final selected architecture is then retrained from scratch. The paper provides no evidence that this weight-sharing proxy correlates with from-scratch performance. If the proxy is noisy, the top-10 set is partly a lottery, and the reported gain may be a selection artifact rather than a genuine architecture advantage. Please supply a correlation analysis between inherited-weight validation mAP and retrained mAP for a random sample of explored architectures, or otherwise justify the reliability of the proxy in this search space.
  3. [Table 1 and §4.2] The claim that SAR-NAS 'outperforms existing SAR detection methods' is based on comparisons with models whose training protocols are unspecified (epochs, batch size, optimizer, augmentation, backbone pretraining, etc.). Some listed baselines (e.g., DETR at 45.73, YOLOF at 42.83) are far below the YOLOv10-N baseline, suggesting possible under-training or different evaluation settings. The abstract and §4.2 present these numbers as SOTA comparison results. Please clarify the training and evaluation protocol for all baselines, or restrict the superiority claim to the controlled YOLOv10 comparison in Table 2 where the experimental setup is matched.
  4. [§3.2 and §4.1] The search space and hardware constraints are not fully specified. §3.2 says each layer's channel configuration is selected from {0.25, 0.5, 0.75, 1.0} of the original width, but it is not stated which backbone layers are searchable (e.g., all 11 layers listed in Table 3?) nor what the exact parameter/FLOPs constraints C are. §4.1 gives P, T, m, p but not the constraint thresholds used to generate SAR-NAS-N and SAR-NAS-S. Without these details, the 'hardware-aware' aspect and the reported cost-efficiency trade-off are not fully reproducible.
minor comments (5)
  1. [§4.3] Typo: 'reducinGFLOPs' should read 'reducing FLOPs'.
  2. [References/§4.2] DenoDet is cited as [19], but reference [19] is DAB-DETR; the correct reference appears to be [8]. Please fix the citation.
  3. [§2.2/References] SpineNet is cited as [1], but [1] is 'Spinnet: Learning a general surface descriptor for 3d point cloud registration', which is a different work. The intended reference is likely Du et al., 'SpineNet: Learning Scale-Permuted Backbone for Recognition and Localization', CVPR 2020.
  4. [Algorithm 1] Variable names are inconsistent: 'T opn', 'T opk', 'P' (population size and probability p) are confusable. Please use distinct and consistently formatted identifiers.
  5. [§4.4] The claim that 'SAR-NAS tends to allocate more channels to early backbone layers' is not quantitatively supported by Table 3 alone; consider adding a summary statistic or a figure showing layer-wise ratios across the top candidates.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: SAR-NAS is a standard one-shot NAS pipeline with independent retraining; no claim reduces to its inputs by construction.

full rationale

The paper's derivation chain is a conventional one-shot NAS loop: train a weight-shared supernet, run an evolutionary search using validation mAP as a proxy (Algorithm 1), retrain the top architectures from scratch, and report the from-scratch mAP. The search-time proxy is explicitly not the final reported result; the top-10 architectures are fully retrained before the final model is selected. The final architecture is therefore not defined as the maximizer of the reported metric, and the reported accuracy is a measured quantity, not a fitted parameter renamed as a prediction. The baseline YOLOv10 is an external input, and the comparison in Table 2 is empirical rather than definitional. There are no load-bearing self-citations: references to sandwich-rule training [34], SPOS [12], SARDet-100K [16], and YOLOv10 [30] are all external works. The only notable weakness is that architecture selection and final evaluation may both use the same validation set, and the paper reports no error bars or repeated searches; this is a statistical overfitting/selection-bias concern, not a circularity of the derivation. The assumed correlation between inherited-weight proxy scores and from-scratch performance is an empirical risk, not a constructional equivalence. Per the rules, absence of a definitional reduction means no circularity is flagged.

Assumptions & free parameters 2 free parameters · 4 assumptions · 0 invented entities

The central claim rests on the fitted channel ratios (free parameters) and on assumptions about the reliability of one-shot NAS evaluation and the benchmark. No new physical entities or quantities are introduced.

free parameters (2)
  • Backbone channel scaling factors = SAR-NAS-S: [1.0,1.0,1.0,1.0,1.0,0.75,1.0,0.5,1.0,0.75,1.0]; SAR-NAS-N: [1.0,1.0,1.0,1.0,1.0,0.75,1.0,0.5,0.75,0.75,1.0]
    11 per-layer channel width ratios selected by evolutionary search to maximize validation mAP; these ratios define the final model and are the primary fitted quantities.
  • Search hyperparameters = P=50, T=20, p=0.1, m=25, top-k=20
    Population size, number of search iterations, mutation/crossover probability, mutation/crossover count, and selection number are chosen by hand without sensitivity analysis; the final result depends on these settings.
assumptions (4)
  • domain assumption SARDet-100K is a valid, representative benchmark for SAR object detection with correct annotations and standard splits.
    Section 4.1 relies on this benchmark for all comparisons and conclusions.
  • domain assumption Weight-shared supernet validation scores from sandwich-rule training accurately rank sub-networks by their true from-scratch potential.
    Sections 3.1 and 3.3: candidates inherit weights and are evaluated on the validation set; the search assumes this ordering transfers to retrained models.
  • domain assumption Validation performance during search predicts test performance for the final architecture.
    Algorithm 1 selects the best architecture based on validation accuracy; the reported test results assume the selection bias is negligible.
  • domain assumption YOLOv10 is a fair baseline and the comparison training recipes in Table 1 are consistent across models.
    The 'state-of-the-art' claim compares models trained under possibly different protocols; the paper does not disclose full training details for all baselines.

how reviews work

0 comments
Cite this review

Pith. "Pith review of SAR-NAS: Lightweight SAR Object Detection with Neural Architecture Search." pith.science (2026). https://pith.science/paper/P53HC4G7

@misc{pith2026250901279,
  author       = {Pith},
  title        = {Pith review of: SAR-NAS: Lightweight SAR Object Detection with Neural Architecture Search},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/P53HC4G7}},
  note         = {Machine review of arXiv:2509.01279}
}
read the original abstract

Synthetic Aperture Radar (SAR) object detection faces significant challenges from speckle noise, small target ambiguities, and on-board computational constraints. While existing approaches predominantly focus on SAR-specific architectural modifications, this paper explores the application of the existing lightweight object detector, i.e., YOLOv10, for SAR object detection and enhances its performance through Neural Architecture Search (NAS). Specifically, we employ NAS to systematically optimize the network structure, especially focusing on the backbone architecture search. By constructing an extensive search space and leveraging evolutionary search, our method identifies a favorable architecture that balances accuracy, parameter efficiency, and computational cost. Notably, this work introduces NAS to SAR object detection for the first time. The experimental results on the large-scale SARDet-100K dataset demonstrate that our optimized model outperforms existing SAR detection methods, achieving superior detection accuracy while maintaining lower computational overhead. We hope this work offers a novel perspective on leveraging NAS for real-world applications.

Figures

Figures reproduced from arXiv: 2509.01279 by the authors.

Figure 1
Figure 1. Comparison of the real-time object detectors on SARDet-100K dataset. SAR-NAS achieves the best accuracy-efficiency trade-off than other models. (1) high false alarm rates and (2) poor generalization across diverse imaging conditions[41]. Recently, deep learning-based object detection methods have dramatically improved performance in SAR object detection. These approaches leverage con￾volutional neural networks (CNNs… view at source ↗
Figure 2
Figure 2. The pipeline of SAR-NAS. 2.2 Neural Architecture Search for Object Detection Neural Architecture Search has emerged as a pivotal technique for automating network design. Early NAS methods relied on reinforcement learning and evolutionary al￾gorithms to iteratively train isolated architectures, which requires high compu￾tational cost. To address this issue, approaches, such as Differentiable Archi￾tecture Search (DAR… view at source ↗
Figure 3
Figure 3. Search space design. 3.1 The Pipeline of SAR-NAS Our SAR-NAS follows One-Shot NAS paradigm with hardware constraints. The search process leverages weight-sharing to inherit pre-trained supernet weights, ensuring efficient and accurate evaluation of candidate architectures. As illus￾trated in [PITH_FULL_IMAGE:figures/full_fig_p005_3.png] view at source ↗

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

42 extracted references · 32 canonical work pages

  1. [1]

    In: CVPR (2021)

    Ao, S., Hu, Q., Yang, B., Markham, A., Guo, Y.: Spinnet: Learning a general surface descriptor for 3d point cloud registration. In: CVPR (2021)

    work page 2021

  2. [2]

    arXiv preprint arXiv:2004.10934 (2020) SAR-NAS 11

    Bochkovskiy, A., Wang, C.Y., Liao, H.Y.M.: Yolov4: Optimal speed and accuracy of object detection. arXiv preprint arXiv:2004.10934 (2020) SAR-NAS 11

    arXiv 2004

  3. [3]

    arXiv preprint arXiv:1908.09791 (2019)

    Cai, H., Gan, C., Wang, T., Zhang, Z., Han, S.: Once-for-all: Train one network and specialize it for efficient deployment. arXiv preprint arXiv:1908.09791 (2019)

    arXiv 1908

  4. [4]

    IEEE TPAMI 43(5) (2021)

    Cai, Z., Vasconcelos, N.: Cascade R-CNN: High quality object detection and instance segmentation. IEEE TPAMI 43(5) (2021). https://doi.org/10.1109/ TPAMI.2019.2956516

    work page arXiv 2021

  5. [5]

    In: ECCV

    Carion, N., Massa, F., Synnaeve, G., Usunier, N., Kirillov, A., Zagoruyko, S.: End- to-end object detection with transformers. In: ECCV. Springer (2020)

    work page 2020

  6. [6]

    In: CVPR (June 2021)

    Chen, Q., Wang, Y., Yang, T., Zhang, X., Cheng, J., Sun, J.: You only look one- level feature. In: CVPR (June 2021)

    work page 2021

  7. [7]

    In: ACM Multimedia (2021)

    Chen, Z., Yang, C., Li, Q., Zhao, F., Zha, Z.J., Wu, F.: Disentangle your dense object detector. In: ACM Multimedia (2021)

    work page 2021

  8. [8]

    IEEE TAES (2024)

    Dai, Y., Zou, M., Li, Y., Li, X., Ni, K., Yang, J.: Denodet: Attention as deformable multi-subspace feature denoising for target detection in sar images. IEEE TAES (2024)

    work page 2024

Show all 42 references
  1. [9]

    In: ICCV (2021)

    Feng, C., Zhong, Y., Gao, Y., Scott, M.R., Huang, W.: TOOD: Task-aligned one- stage object detection. In: ICCV (2021)

    2021

  2. [10]

    arXiv preprint arXiv:2107.08430 (2021)

    Ge, Z., Liu, S., Wang, F., Li, Z., Sun, J.: Yolox: Exceeding yolo series in 2021. arXiv preprint arXiv:2107.08430 (2021)

    2021 arXiv

  3. [11]

    In: CVPR (2019)

    Ghiasi, G., Lin, T.Y., Le, Q.V.: Nas-fpn: Learning scalable feature pyramid archi- tecture for object detection. In: CVPR (2019)

    2019

  4. [12]

    In: ECCV

    Guo, Z., Zhang, X., Mu, H., Heng, W., Liu, Z., Wei, Y., Sun, J.: Single path one-shot neural architecture search with uniform sampling. In: ECCV. Springer (2020)

    2020

  5. [13]

    In: ECCV

    Kim, K., Lee, H.S.: Probabilistic anchor assignment with IoU prediction for object detection. In: ECCV. Springer (2020)

    2020

  6. [14]

    In: NeurIPS (2020)

    Li, X., Wang, W., Wu, L., Chen, S., Hu, X., Li, J., Tang, J., Yang, J.: General- ized focal loss: Learning qualified and distributed bounding boxes for dense object detection. In: NeurIPS (2020)

    2020

  7. [15]

    In: IGARSS

    Li, Y., Ding, Z., Zhang, C., Wang, Y., Chen, J.: Sar ship detection based on resnet and transfer learning. In: IGARSS. IEEE (2019)

    2019

  8. [16]

    arXiv preprint arXiv:2403.06534 (2024)

    Li, Y., Li, X., Li, W., Hou, Q., Liu, L., Cheng, M.M., Yang, J.: Sardet-100k: Towards open-source benchmark and toolkit for large-scale sar object detection. arXiv preprint arXiv:2403.06534 (2024)

    2024 arXiv

  9. [17]

    In: ICCV (Oct 2017)

    Lin, T.Y., Goyal, P., Girshick, R., He, K., Dollar, P.: Focal loss for dense object detection. In: ICCV (Oct 2017)

    2017

  10. [18]

    arXiv preprint arXiv:1806.09055 (2018)

    Liu, H., Simonyan, K., Yang, Y.: Darts: Differentiable architecture search. arXiv preprint arXiv:1806.09055 (2018)

    2018 arXiv

  11. [19]

    In: ICCV (2022)

    Liu, S., Li, F., Zhang, H., Yang, X., Qi, X., Su, H., Zhu, J., Zhang, L.: DAB-DETR: Dynamic anchor boxes are better queries for DETR. In: ICCV (2022)

    2022

  12. [20]

    IEEE TGRS59(5) (2020)

    Liu, S., Gao, L., Lei, Y., Wang, M., Hu, Q., Ma, X., Zhang, Y.D.: Sar speckle removal using hybrid frequency modulations. IEEE TGRS59(5) (2020)

    2020

  13. [21]

    IEEE JSTARS12(10) (2019)

    Liu, T., Yang, Z., Yang, J., Gao, G.: Cfar ship detection methods using compact polarimetric sar in a k-wishart distribution. IEEE JSTARS12(10) (2019)

    2019

  14. [22]

    In: CVPR (2022)

    Liu, Z., Mao, H., Wu, C.Y., Feichtenhofer, C., Darrell, T., Xie, S.: A convnet for the 2020s. In: CVPR (2022)

    2022

  15. [23]

    Lu,W.,Fan,Z.,Wei,L.,Xiao-ming,X.,Wei,H.:Amethodofsartargetrecognition based on gabor filter and local texture feature extraction. J. Radars4(6) (2015)

    2015

  16. [24]

    In: CVPR (June 2019)

    Pang, J., Chen, K., Shi, J., Feng, H., Ouyang, W., Lin, D.: Libra R-CNN: Towards balanced learning for object detection. In: CVPR (June 2019)

    2019

  17. [25]

    NeurIPS28 (2015) 12 X

    Ren, S., He, K., Girshick, R., Sun, J.: Faster R-CNN: Towards real-time object detection with region proposal networks. NeurIPS28 (2015) 12 X. Yu et al

    2015

  18. [26]

    In: BIGSARDATA

    Shao, S., Li, H., Wang, S.: Sar ship detection from complex background based on dynamic shrinkage attention mechanism. In: BIGSARDATA. IEEE (2021)

    2021

  19. [27]

    In: CVPR (2020)

    Tan, M., Pang, R., Le, Q.V.: Efficientdet: Scalable and efficient object detection. In: CVPR (2020)

    2020

  20. [28]

    In: ICCV (October 2019)

    Tian, Z., Shen, C., Chen, H., He, T.: FCOS: Fully convolutional one-stage object detection. In: ICCV (October 2019)

    2019

  21. [29]

    IEEE TGRS60 (2021)

    Wan, H., Chen, J., Huang, Z., Xia, R., Wu, B., Sun, L., Yao, B., Liu, X., Xing, M.: Afsar: An anchor-free sar target detection algorithm based on multiscale en- hancement representation learning. IEEE TGRS60 (2021)

    2021

  22. [30]

    NeurIPS37 (2024)

    Wang, A., Chen, H., Liu, L., Chen, K., Lin, Z., Han, J., et al.: Yolov10: Real-time end-to-end object detection. NeurIPS37 (2024)

    2024

  23. [31]

    In: ICCV (2021)

    Wang, W., Xie, E., Li, X., Fan, D.P., Song, K., Liang, D., Lu, T., Luo, P., Shao, L.: Pyramid vision transformer: A versatile backbone for dense prediction without convolutions. In: ICCV (2021)

    2021

  24. [32]

    IEEE JSTARS14 (2021)

    Wang, Z., Du, L., Li, Y.: Boosting lightweight cnns through network pruning and knowledge distillation for sar target recognition. IEEE JSTARS14 (2021)

    2021

  25. [33]

    NeurIPS34 (2021)

    Wu, H., Xu, J., Wang, J., Long, M.: Autoformer: Decomposition transformers with auto-correlation for long-term series forecasting. NeurIPS34 (2021)

    2021

  26. [34]

    arXiv preprint arXiv:1812.08928 (2018)

    Yu, J., Yang, L., Xu, N., Yang, J., Huang, T.: Slimmable neural networks. arXiv preprint arXiv:1812.08928 (2018)

    2018 arXiv

  27. [35]

    In: CVPR (June 2021)

    Zhang, H., Wang, Y., Dayoub, F., Sunderhauf, N.: VarifocalNet: An IoU-aware dense object detector. In: CVPR (June 2021)

    2021

  28. [36]

    In: ECCV

    Zhang, H., Chang, H., Ma, B., Wang, N., Chen, X.: Dynamic R-CNN: Towards high quality object detection via dynamic training. In: ECCV. Springer (2020)

    2020

  29. [37]

    In: CVPR (June 2020)

    Zhang, S., Chi, C., Yao, Y., Lei, Z., Li, S.Z.: Bridging the gap between anchor- based and anchor-free detection via adaptive training sample selection. In: CVPR (June 2020)

    2020

  30. [38]

    Remote Sensing11(10) (2019)

    Zhang, T., Zhang, X.: High-speed ship detection in sar images based on a grid convolutional neural network. Remote Sensing11(10) (2019)

    2019

  31. [39]

    arXiv preprint arXiv:1904.07850 (2019)

    Zhou, X., Wang, D., Krähenbühl, P.: Objects as points. arXiv preprint arXiv:1904.07850 (2019)

    1904 arXiv

  32. [40]

    arXiv preprint arXiv:2007.03496 (2020)

    Zhu, B., Wang, J., Jiang, Z., Zong, F., Liu, S., Li, Z., Sun, J.: Autoassign: Differen- tiable label assignment for dense object detection. arXiv preprint arXiv:2007.03496 (2020)

    2007 arXiv

  33. [41]

    IEEE GRSM 9(4) (2021)

    Zhu, X.X., Montazeri, S., Ali, M., Hua, Y., Wang, Y., Mou, L., Shi, Y., Xu, F., Bamler, R.: Deep learning meets sar: Concepts, models, pitfalls, and perspectives. IEEE GRSM 9(4) (2021)

    2021

  34. [42]

    In: ICLR (2021)

    Zhu, X., Su, W., Lu, L., Li, B., Wang, X., Dai, J.: Deformable DETR: Deformable transformers for end-to-end object detection. In: ICLR (2021)

    2021

Pith tools

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