PaperlayerView exact revision on arXiv

Source-linked AI summary

Strip Convolution and Direction-Aware Exclusion Loss for Oriented Ship Detection

Bin Chen, Yuanyuan Liu, Peng Yang, Chao Lu

arXiv:2609.27238v1cs.CV

TL;DR

Oriented ship detection must handle elongated hull geometry and dense port scenes, while existing approaches typically treat representation and duplicate suppression separately. The paper combines orthogonal strip convolutions with a class-, direction-, and confidence-aware exclusion loss. It reports 78.45% mAP50:95 on HRSC2016 and 53.71% on DIOR-R with 2.91M parameters.

  • Problem

    Elongated hulls and dense ship distributions challenge oriented detection, while existing methods typically address geometric representation and duplicate suppression separately.

    Source: §I · p. 1

  • Method

    The YOLOv11-OBB-based detector combines orthogonal strip convolutions for elongated hulls with CA-DAEL, which uses class, orientation, overlap, and confidence cues to suppress redundant predictions.

    Source: §I · pp. 1–2

  • Results

    78.45% mAP50:95 on HRSC2016 and 53.71% mAP50:95 on DIOR-R are achieved with 2.91M parameters.

    Source: §I · p. 2

  • Takeaways & Limitations

    The reported results support the method’s effectiveness for oriented ship detection on HRSC2016 and DIOR-R.

    Source: §I · p. 2

Abstract

from arXiv · show

Oriented ship detection in very high resolution (VHR) remote sensing imagery remains challenging due to elongated hull geometry and dense target distributions in complex port scenes. Existing methods typically address geometric representation and duplicate suppression separately. To jointly tackle these issues, we propose an oriented ship detector with two complementary components. The C3k2_Strip module employs orthogonal strip convolutions to better capture elongated hull structures, while the Class-Aware Direction-Aware Exclusion Loss (CA-DAEL) suppresses redundant predictions using class, direction, and confidence cues. Experiments on HRSC2016 and DIOR-R achieve 78.45% and 53.71% mAP50:95, respectively, with only 2.91M parameters. On HRSC2016, the proposed method improves mAP50:95 by 6.32 percentage points over the YOLOv11-OBB baseline, demonstrating its effectiveness for accurate oriented ship detection.

I. INTRODUCTION

The introduction identifies elongated hulls and dense dockside overlap as distinct but coupled obstacles to oriented ship detection. It proposes combining strip convolutions with a class-, direction-, and confidence-aware exclusion loss, and reports results on two benchmarks.

I. INTRODUCTION

  • Elongated hulls can be poorly represented by isotropic receptive fields, weakening directional responses for long, narrow vessels.The passage notes that hull aspect ratios can reach 15:1 and describes additional computational or sequential overhead in deformable and snake convolutions.Source: §I · p. 1
  • Overlapping OBBs between neighboring ships create localization ambiguity and redundant predictions that can bias regression and weaken NMS reliability.The introduction argues that an OBB-specific exclusion mechanism should suppress duplicates without repelling legitimate adjacent ships.Source: §I · p. 1
  • C3k2 Strip and CA-DAEL jointly address elongated-hull representation and dense-scene duplicates in a YOLOv11-OBB-based detector.Strip convolutions extend receptive fields horizontally and vertically; CA-DAEL uses class, orientation, overlap, and confidence cues to suppress redundant same-class predictions while preserving valid neighbors.Source: §I · pp. 1–2
  • The two components are presented as complementary: strip-aligned features improve representation, while CA-DAEL reduces prediction redundancy during training.The method is described as addressing both challenges jointly.Source: §I · p. 1
  • 78.45% and 53.71% mAP50:95 on HRSC2016 and DIOR-R, respectively, are achieved with 2.91M parameters.The introduction presents these experiments as evidence of the method’s effectiveness and generalization.Source: §I · p. 2

II. PROPOSED METHOD

The proposed pipeline inherits a rotated-YOLO single-stage anchor-free framework and adds C3k2 Strip modules and CA-DAEL. Its OBB heads predict rotated boxes using the long-edge angle convention.

Overall Framework

  • The rotated-YOLO pipeline uses a CSPDarknet backbone and OBB heads inherited from YOLOv11-OBB, with C3k2 Strip modules and CA-DAEL as the novel components.The backbone extracts multi-scale features P3–P5; the heads output rotated boxes.Source: §II · p. 2
  • The OBB heads output (x, y, w, h, θ) boxes, with θ in [0°, 180°) under the long-edge convention.Source: §II · p. 2
  • The model comprises 2.91M parameters and 7.25 GFLOPs.Source: §II · p. 2

A. C3k2 Strip Module for Elongated Features

C3k2 Strip augments standard bottlenecks with orthogonal strip-convolution attention to strengthen hull-aligned features. Its fixed strip length is matched to vessel scale, with a reported 4.41-point mAP50:95 gain over the baseline for an 8.4% FLOPs overhead.

A. C3k2 Strip Module for Elongated Features

  • The module retains two 3×3 convolutions and adds a strip-convolution attention block that selectively enhances hull-aligned responses.The 3×3 convolutions provide general feature transformation, while the strip block serves as a complementary attention prior.Source: A. C3k2 Strip Module for Elongated… · p. 2
  • A 5×5 depthwise step expands local context along both axes before horizontal and vertical strip convolutions capture hull-aligned features.The orthogonal kernels use k1×k2 and k2×k1 shapes.Source: A. C3k2 Strip Module for Elongated… · p. 2
  • Strip features form an attention map that modulates the bottleneck feature, and the residual sum forms each bottleneck output within the CSP path.The CSP path splits the input into a bypass branch and a processed branch, then fuses the branches with a final 1×1 convolution.Source: A. C3k2 Strip Module for Elongated… · p. 2, Figure 3 · p. 2
  • An 8.4% FLOPs overhead yields +4.41 mAP50:95 for C3k2 Strip over the baseline (Table III, R1→R2).The model uses 2.91M parameters and 7.25 GFLOPs, versus 2.66M parameters and 6.69 GFLOPs for the baseline.Source: A. C3k2 Strip Module for Elongated… · p. 2
  • A fixed strip length k2 = 7 spans 56, 112, and 224 input pixels at neck strides 8, 16, and 32, respectively.The paper relates this scale to a roughly 200-pixel 300-m tanker at 1.5-m GSD, fully covered at P5.Source: A. C3k2 Strip Module for Elongated… · p. 2

B. Class-Aware Direction-Aware Exclusion Loss

CA-DAEL augments standard rotated-YOLO losses with a gated exclusion term to address redundant overlapping predictions in dense port scenes. Its gates target same-class, highly overlapping pairs while preserving neighboring vessels.

B. Class-Aware Direction-Aware Exclusion Loss

  • CA-DAEL adds a class-, direction-, and confidence-gated exclusion term to standard rotated-YOLO losses to address redundant predictions that conventional losses and NMS may not resolve.The standard loss also includes rotated CIoU regression, BCE classification, Distribution Focal Loss for box-edge refinement, and the baseline angle loss.Source: B. Class-Aware Direction-Aware… · p. 3
  • Pairs of same-class predictions are penalized above τ=0.8 IoU, with λexc=0.3 and a 100-epoch loss-specific warmup.The warmup is separate from the 3-epoch learning-rate warmup; the high threshold is intended to avoid penalizing legitimate adjacent vessels, whose inter-OBB IoU typically stays in the 0.6–0.7 range.Source: B. Class-Aware Direction-Aware… · p. 3, Figure 4 · p. 3
  • The direction factor strengthens penalties for parallel pairs, while the class indicator prevents cross-class interference and lower-overlap detections on adjacent vessels are preserved.The direction factor is most informative when predicted angles align with the true hull axis, motivating its pairing with C3k2 Strip features.Source: B. Class-Aware Direction-Aware… · p. 3, Figure 4 · p. 3

A. Datasets and Evaluation Protocol

The evaluation uses HRSC2016 and DIOR-R, with polygon-based rotated IoU following the MMRotate protocol.

A. Datasets and Evaluation Protocol

B. Implementation Details

Experiments use YOLOv11-OBB as the baseline, with a separate CA-DAEL warmup and specified training settings for each dataset.

B. Implementation Details

  • Experiments use the Ultralytics framework with YOLOv11-OBB as the baseline, on an Intel Xeon E5-2680 v4 system with 64 GB RAM and an NVIDIA RTX 3080 Ti.The software environment is Python 3.10.20, PyTorch 2.5.1, and CUDA 12.1.Source: B. Implementation Details · p. 3
  • Models use 640 × 640 inputs, AdamW with LR 0.002, momentum 0.937, weight decay 5 × 10−4, linear LR scheduling, and batch size 32.Training lasts 200 epochs on HRSC2016 and 300 epochs on DIOR-R, with a 3-epoch learning-rate warmup.Source: B. Implementation Details · p. 3
  • Augmentation includes ±30° rotation, ±10° shear, and Mosaic disabled for the final 10 epochs; loss weights are tuned on the HRSC2016 validation set.A separate 100-epoch CA-DAEL warmup delays the exclusion term until regression and classification stabilize.Source: B. Implementation Details · p. 3

C. Comparison with State-of-the-Art Methods

Under matched training protocols, the proposed method leads the HRSC2016 comparison, while the efficiency frontier places it at the highest accuracy among models under 3M parameters.

C. Comparison with State-of-the-Art Methods

  • All methods in Tables I–II were re-implemented with identical training and evaluation protocols, including 200 HRSC2016 epochs and 300 DIOR-R epochs at 640 × 640 input resolution.Source: C. Comparison with State-of-the-Art… · p. 3
  • The proposed model is Pareto-optimal on HRSC2016, with the highest accuracy under 3M parameters; generic non-YOLO detectors exceed 4M parameters and remain below 47% mAP50:95.Source: Figure 5 · p. 4

D. Cross-Dataset Generalization on DIOR-R

On DIOR-R, the proposed method improves over YOLOv11-OBB, and the reported gains extend to heterogeneous elongated targets and categories with frequent spatial overlap.

D. Cross-Dataset Generalization on DIOR-R

  • 71.47% mAP50 and 53.71% mAP50:95 on DIOR-R improve over YOLOv11-OBB by 5.00 and 3.37 points, respectively.The reported scores are the 20-class mean.Source: D. Cross-Dataset Generalization on… · p. 4
  • The qualitative comparison uses two HRSC2016 dense-dock scenes: one with 8 ships and one with 6 ships in tight parallel mooring.Source: Figure 6 · p. 4
  • C3k2 Strip generalizes to heterogeneous elongated targets such as vehicles and aircraft, while CA-DAEL remains advantageous on categories with frequent spatial overlap.Source: D. Cross-Dataset Generalization on… · p. 4

E. Ablation Studies

The combined detector reaches 78.45% mAP50:95 on HRSC2016 and 53.71% on DIOR-R; ablations and dense-dock examples report complementary contributions from strip alignment and exclusion gating.

E. Ablation Studies

  • C3k2 Strip alone adds +4.41 mAP50:95 points for 0.25M parameters and 0.56 GFLOPs, while CA-DAEL alone adds +3.65 points with zero parameter overhead; together they reach 78.45% (+6.32).The standalone gains sum to 8.06, leaving 1.74 mAP50:95 points of overlap in the combined gain.Source: E. Ablation Studies · p. 4

F. Qualitative Analysis

  • In the 8-ship scene, false alarms fall from 14 to 4; in the 6-ship scene, the proposed method has zero false alarms and correctly bounds each ship.The comparison describes baseline angle deviations and redundant boxes in the 8-ship scene, alongside predictions from orthogonal strip kernels and CA-DAEL gating.Source: F. Qualitative Analysis · p. 5

IV. CONCLUSION

  • The 2.91M-parameter model reaches 78.45% mAP50:95 on HRSC2016 and 53.71% on DIOR-R, surpassing YOLOv8-OBB by 4.37 points at 7.25 GFLOPs.Future work will explore aspect-ratio-adaptive strip kernels and multi-seed validation across diverse oriented object detection scenarios.Source: §IV · p. 5

Related papers

Loading 2609.27238v1…