acceptodds
Under review as a conference paper at ICLR 2027

SARoLO: Structure and Scale-Aware SAR Object Detection

Abstract

Synthetic Aperture Radar (SAR) object detection faces persistent challenges caused by small-target dominance, strong speckle noise, and complex background clutter that obscure fine-grained structural cues. Although conventional YOLO detectors excel in general-purpose visual tasks with balanced accuracy and efficiency, they remain limited in representing scale- and structure-sensitive features due to their reliance on aggressive downsampling, information-lossy bottleneck compression, and redundant multi-scale detection heads. To address these limitations, we propose a novel structure and scale-aware SAR object detection method, called SARoLO, including a Heterogeneous Triple-Conv Downsampling (HetTriDown) module, a Heterogeneous Convolution Modulation Block (HCM Block) with dynamic nonlinear modulation (DyN), a cross-stage feature aggregation module (HCMC3K), a Cross-Scale Aggregation Network (CSAN), and a dual-scale detection head, to optimize bounding box performance for SAR imagery characterized by small/medium targets and significant structural variations. Extensive experiments on the HRSID and SSDD benchmarks demonstrate SARoLO's state-of-the-art performance. On HRSID, SARoLO achieves 68.46% AP, a 2.52% gain over the latest YOLOv13 baseline. On SSDD, SARoLO attains 63.21% AP, a 0.9% improvement over YOLOv13. These results validate its effectiveness in structure- and scale-aware feature learning and efficient SAR object detection.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
SAR, object detection, YOLO

Reviews

Among:
Shape:
1234this paper, Bottom 1%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.00 over 3 reviews, above 0% of the others.
Reviewer_QdxVOpenReview ↗
Rating
1
Confidence
4
Informativeness
1
Presentation
2
Soundness
1
Read review
Paper summary

This paper proposes SARoLO, a modification of YOLO11 for SAR object detection. It introduces alternative downsampling and bottleneck blocks, cross-scale feature fusion, and a dual-scale detection head. The design combines square and asymmetric convolutions with a learnable nonlinear transformation called DyN. Results are reported on HRSID, SSDD, and, in the supplementary material, RSAR.

Critical strengths

Detecting small ships in cluttered SAR imagery is a relevant application. The proposed model has a small parameter count and reports accuracy improvements over the tested YOLO baselines.

Critical weaknesses
  1. The attribution and novelty of central components are not adequately established. Eq. (1), introduced as DyN, is identical to the full DyT operation in Transformers without Normalization, including the learnable input scaling and per-channel affine parameters. This work is neither cited nor acknowledged. Since DyN is repeatedly presented as a component designed for SARoLO, this directly affects the originality assessment.
  2. The combination of square and asymmetric convolutions also needs to be discussed against closely related work, including ACNet, Strip R-CNN, and PKINet-v2. Their implementations are not identical to SARoLO, but they are directly relevant to its architectural choices and motivation concerning elongated targets.
  3. The experiments do not sufficiently separate the proposed mechanisms from simpler architectural changes and additional computation. Table 3 shows that changing YOLO11 to two detection heads already improves AP from 66.36 to 67.10. The complete SARoLO reaches 68.46, but increases computation from 5.8G to 10.0G FLOPs compared with this two-head baseline, while reducing parameters only slightly, from 1.85M to 1.80M. Without measured inference latency, the smaller parameter count does not establish an overall efficiency improvement.
Questions for authors
  1. Is there any difference between DyN in Eq. (1) and DyT in Transformers without Normalization? If so, where is that difference? Please clarify the attribution and which part should be considered a new contribution.
  2. How do the proposed square-and-asymmetric convolution blocks differ from the relevant designs in ACNet, Strip R-CNN, and PKINet-v2? Do the authors have comparisons of these components under the same SAR detection setting?
  3. How does SARoLO compare with the two-head YOLO11 baseline at a similar computational budget? Measured latency and a capacity-matched replacement for CSAN would help determine whether the proposed design offers a better accuracy–efficiency trade-off.
Reviewer_qM8pOpenReview ↗
Rating
1
Confidence
4
Informativeness
1
Presentation
2
Soundness
1
Read review
Paper summary

The paper proposes SARoLO to improve detection of small and medium-sized SAR targets whose details are obscured by speckle noise and background clutter. It modifies YOLO11 using asymmetric convolution branches, learnable activations, revised downsampling and feature fusion, and two detection heads. Experimental results show the effectiveness of the proposed method.

Critical strengths

Experimental results show the effectiveness of the proposed method.

Critical weaknesses

I would like to flag potential plagiarism, unattributed reuse and citation integrity concerns:

The proposed DyN in Eq. (1) is identical to DyT in Transformers without Normalization [1], which is neither cited nor acknowledged. Highly relevant prior work—ACNet (ICCV 2019), Strip R-CNN (AAAI 2026), and PKINet-v2 (preprint)—is omitted. The latter two share the paper’s central motivation concerning anisotropic structures and elongated remote-sensing targets. Line 86 attributes “FPN, PAN, and BiFPN” to Lei et al. (2025), but that reference is actually YOLOv13. Together, these raise substantial concerns about contribution attribution and the accuracy of the novelty framing.

[1] Zhu, J., Chen, X., He, K., LeCun, Y. and Liu, Z., 2025, June. Transformers without normalization. In 2025 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) (pp. 14901-14911). IEEE.

Questions for authors

See above

Reviewer_RmJVOpenReview ↗
Rating
1
Confidence
4
Informativeness
2
Presentation
3
Soundness
1
Read review
Paper summary

This paper has made some improvements based on YOLO11. This paper proposes Heterogeneous Triple-Conv Down sampling, learnable tanh modulation, across-stage feature aggregation module, and dual scale detection heads. These modules can improve the performance of SAR target detection. However, the ablation of some key modules is insufficient and there is a risk of overlap with other previous works.

Critical strengths

1.SAR target detection is a valuable issue. The modules introduced in the paper can improve the performance of SAR target detection. 2.The ablation experiment in the paper covers the introduced modules.

Critical weaknesses

1.The core innovation of this paper is insufficient. The formula for DyN proposed in the paper is consistent with CVPR2025's "Transformers without Normalization", but this work is not cited in this paper. This module may simply be an application of the previously proposed Dynamic Tanh to SAR target detection. The HetTriDown module used in this paper follows Adown's (proposed in YOLOV9) average pooling, channel branching, down sampling, and concatenation module. This work makes only an incremental contribution. 2. Experimental comparison is unfair. According to the supplementary materials, YOLO series methods include SARoLO using 300 epochs of training and data augmentation methods such as Mosaic. The official baselines for RSAR and DETR series methods use 12 or 36 epochs and different augmentation methods. At least the same data augmentation methods should be used for fair comparison.

Questions for authors
  1. What is the difference between DyN and DyT?
  2. Does HetDriDown have performance advantages compared to Adown (YOLOV9)?

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.