PaperlayerView exact revision on arXiv

Source-linked AI summary

KSG-Net: Key-Sparse and Global-Context Learning for Maritime 3D Ship Detection

Zhouyuan Huai, Meiqi Wan, Yan Yang, Minshi Chen, Xin Yuan, Wei Wang, Xiao Wang

arXiv:2609.02077v1cs.CV

TL;DR

Maritime 3D ship detection must cope with scale variation, sparse small-vessel observations, clutter, and limited global modeling. KSG-Net combines key-sparse multi-scale aggregation with global-context aggregation in a fully sparse framework, achieving higher mAP than compared methods across real and simulated Thames River datasets.

  • Problem

    Maritime LiDAR detection lacks a unified fully sparse approach for stable small-vessel representation and global structural modeling of large vessels.

  • Method

    KSG-Net jointly uses KSMA for key-voxel multi-scale aggregation and GCA for scene-level global context with gated residual interactions.

  • Results

    87.89% mAP on Thames River Real and 84.44% on Thames River Simulated, with the best overall mAP on both datasets.

  • Takeaways & Limitations

    KSG-Net provides a fully sparse maritime detector that improves multi-scale vessel detection and robustness in complex maritime environments.

  • Takeaways & Limitations

    Existing sparse detectors remain limited by their autonomous-driving design, while maritime detection methods commonly rely on nonfully-sparse paradigms.

Abstract

from arXiv · show

Accurate 3D ship detection in maritime environments is critical for autonomous navigation, yet remains challenging due to large-scale vessel variations, sparse point clouds of small vessels, and severe sea-clutter interference. Existing methods, primarily based on 2D features or dense representations, struggle to balance detection accuracy and computational efficiency, while sparse 3D detectors designed for road scenes generalize poorly to maritime scenarios. This paper focuses on two key challenges in maritime LiDAR perception: weak feature representation for small and sparse vessels, and insufficient global structural modeling for large vessels due to the limited receptive field of local sparse convolutions. To address these issues, we propose KSG-Net, a Key-Sparse and Global-Context learning network for maritime 3D ship detection. The core idea is to jointly enhance local discriminative features and global structural awareness within a unified fully sparse detection framework. Specifically, a Key Sparse Multi-scale Aggregation (KSMA) module is designed to enhance the representation of small and sparse vessels by selecting informative key voxels and aggregating cross-scale neighborhood features. Furthermore, a Global Context Aggregation (GCA) module is introduced to capture long-range geometric dependencies through scene-level context modeling with gated residual interactions, thereby improving the representation of large vessels. Extensive experiments on the Thames River vessel dataset and simulated datasets demonstrate that KSG-Net consistently outperforms existing methods in multi-scale vessel detection and exhibits strong robustness in complex maritime environments.

1 Introduction

Maritime 3D ship detection must handle scale variation, sparse observations, irregular backgrounds, and sea clutter. KSG-Net addresses weak small-vessel features and limited large-vessel global modeling through key-sparse aggregation and global context within a unified sparse framework.

  • 1 Introduction: Maritime 3D ship detection is important for autonomous navigation, collision avoidance, and intelligent decision-making under complex sensing conditions.The environment includes large vessel-scale variations, irregular backgrounds, sparse point observations, and severe sea-clutter interference.
  • 1 Introduction: Large vessels span broad regions, but local sparse convolutions have limited receptive fields for capturing long-range structural dependencies.
  • 1 Introduction: Small vessels generate sparse, discontinuous point responses that provide limited geometric cues and can be confused with background clutter.
  • 1 Introduction: KSG-Net jointly enhances local discriminative representations and global structural awareness within a unified sparse detection framework.
  • 1 Introduction: KSMA selects informative key voxels and aggregates cross-scale neighborhoods, while GCA models scene-level context and injects it through gated residual interactions.
  • 1 Introduction: Experiments on Thames River and simulated datasets report stronger multi-scale vessel detection and robustness in complex maritime environments than existing methods.

2 Related Work

Prior maritime and sparse 3D detectors face efficiency, geometric-preservation, domain-transfer, and global-modeling limitations. KSG-Net builds on fully sparse detection to jointly target small sparse vessels and large-vessel structure in maritime LiDAR scenes.

  • 2 Related Work: Dense detectors use structured representations such as BEV feature maps, but densification introduces quantization errors and computational redundancy.
  • 2 Related Work: Dense methods can also struggle to preserve fine-grained geometric structures of small and sparse vessels in maritime LiDAR scenarios.
  • 2 Related Work: Sparse detectors improve efficiency by extracting features and predictions only on non-empty voxels or foreground points.
  • 2 Related Work: Fully sparse frameworks and long-range modeling methods have advanced sparse detection through unified sparse heads, virtual voxels, and hierarchical interactions.
  • 2 Related Work: Most sparse detectors target autonomous driving, whose structured layouts and controlled backgrounds differ from maritime scenes with irregular distributions and stronger interference.
  • 2 Related Work: Maritime perception methods commonly rely on images, fusion, dense representations, or nonfully-sparse point-cloud paradigms, leaving unified fully sparse detection underexplored.
  • 2 Related Work: KSG-Net starts from a fully sparse framework and uses KSMA and GCA to enhance key sparse representation and global context modeling, respectively.

3 Methodology

KSG-Net is a fully sparse maritime 3D detection framework that sequentially strengthens local features for small vessels and global structural modeling for large vessels. KSMA selects informative voxels and aggregates multi-scale neighborhoods, while GCA injects gated scene-level context before prediction.

  • Overall Architecture: KSG-Net uses input encoding, sparse backbone extraction, KSMA, GCA, and a detection head in a fully sparse pipeline.The modules operate sequentially on a shared sparse BEV representation.
  • Key Sparse Multi-scale Aggregation: KSMA identifies informative key voxels, aggregates cross-scale sparse neighborhoods, and writes refined responses back through adaptive fusion and residual enhancement.This selecting-before-enhancing strategy concentrates computation on informative sparse locations.
  • Key Sparse Multi-scale Aggregation: A sparse scoring head assigns saliency scores to non-empty voxels, after which the top-K voxels in each sample anchor subsequent aggregation.Saliency-aware selection avoids uniformly enhancing all non-empty voxels and reduces background-clutter interference.
  • Key Sparse Multi-scale Aggregation: KSMA mean-aggregates neighborhood features at multiple scales, projects them separately, predicts adaptive scale weights, and fuses the resulting representations.The scale-weight vector is normalized and varies across key voxels.
  • Key Sparse Multi-scale Aggregation: KSMA concatenates each original key feature with its fused context and uses saliency to gate the residual update magnitude.A learnable residual scaling coefficient controls the feature-fusion update.
  • Global Context Aggregation: GCA globally pools non-empty sparse locations, projects the scene descriptor, broadcasts it to voxels, and injects it through gated residual interactions.This lightweight context modeling supports long-range dependencies without explicit pairwise global self-attention.
  • Detection Head: The GCA-enhanced sparse features feed a detection head predicting classification, offsets, box dimensions, orientation, and an IoU-related branch.The model is trained end-to-end within a unified fully sparse detection framework.

4 Experiments

Experiments show that KSG-Net delivers strong multi-scale maritime detection across real, simulated, and Ship LiDAR datasets while retaining sparse-processing efficiency. Ablations and hyperparameter studies support complementary KSMA and GCA contributions and stable settings across datasets.

  • Ship LiDAR Dataset: KSG-Net achieves 85.65% mAP@0.5 on the Ship LiDAR Dataset, exceeding the strongest BEV and voxel baselines.It surpasses Fade3D by 6.40%, SAFDNet by 7.58%, and DSVT by 8.21%.
  • Ship LiDAR Dataset: KSG-Net uses 31.42 G FLOPs and runs at 33.11 FPS, remaining comparable to VoxelNeXt and GBlobs while costing less than DSVT and AutoReg3D.
  • Thames River Datasets: KSG-Net is strongest for heterogeneous or sparse vessel categories, while competitors retain isolated advantages on Cargo, Large-vessel, or Small-vessel AP.It achieves the best Tour boat, Engineering ship, and Speedboat AP on Ship LiDAR, and the best Large-vessel AP on Thames Simulated.
  • Thames River Datasets: KSG-Net obtains the best mAP on both Thames datasets: 87.89% on Real and 84.44% on Simulated.On Thames Real, it exceeds GBlobs by 4.57% and DSVT by 6.20%; on Thames Simulated, it exceeds GBlobs, LION, and Fade3D.
  • Qualitative Analysis: Feature visualizations and qualitative comparisons show focused sparse-vessel responses, coherent hull activations, and fewer missed or erroneous detections.The qualitative evidence covers vessels at different scales in complex real maritime scenes.
  • Ablation Study: The full framework improves mAP from 78.74% to 87.89% on Thames Real and from 77.13% to 84.44% on Thames Simulated.These correspond to gains of +9.15% and +7.31%, respectively, over the ablation baseline.
  • Ablation Study: KSMA and GCA provide complementary gains: KSMA improves sparse-target categories, GCA raises overall mAP, and their combination performs best.KSMA improves Small-vessel AP by 17.37% on Thames Real, while GCA raises overall mAP by 2.18% and 2.36% on Real and Simulated datasets.
  • Hyperparameter Analysis: The best performance on both Thames datasets uses voxel size [0.2, 0.2, 0.4], a TopK ratio of 15%, and eight KNN neighbors.The study varies voxel size, key-voxel selection ratio, and neighborhood size.

5 Conclusion

KSG-Net addresses maritime 3D ship detection by jointly strengthening local representations for small vessels and global structural awareness for large vessels in a fully sparse framework. Experiments on real and simulated Thames River datasets show consistent multi-scale detection improvements and robustness in complex maritime environments.

  • KSG-Net targets vessel detection under sparse observations, varying scales, and severe background interference in maritime scenes.
  • The network jointly enhances local discriminative representations for small vessels and global structural awareness for large vessels within a unified fully sparse detector.
  • KSMA strengthens small and sparse vessels through informative key-voxel selection and cross-scale neighborhood aggregation.
  • GCA captures long-range geometric dependencies for large vessels using scene-level global-context modeling with gated residual interactions.
  • Experiments on Thames River vessel and simulated datasets show that KSG-Net consistently outperforms existing methods in multi-scale vessel detection and remains robust in complex maritime environments.
Loading 2609.02077v1…