SOAM Block: A Scale-Orientation-Aware Module for Efficient Object Detection in Remote Sensing Imagery

Yi Chen, Zhidong Wang, Zhipeng Xiong, Yufeng Zhang, Xinqi XuSymmetry 2025 · DOI

TL;DR

This paper introduces the SOAM Block, a novel module for object detection in remote sensing imagery that explicitly models the coupling between object scale and orientation. By integrating multi-scale feature extraction with a context-aware attention mechanism and an alignment convolution, SOAM Block achieves state-of-the-art performance on challenging datasets while maintaining computational efficiency.

What Problem Does This Paper Address?

Object detection in remote sensing imagery is crucial for applications like environmental monitoring and urban planning. However, it's a challenging task due to significant variations in object scale, arbitrary orientations, and complex backgrounds. Existing methods often struggle to effectively handle these challenges because they tend to model scale and orientation independently, overlooking their intrinsic relationship. This can lead to issues like fixed receptive fields failing to capture sufficient contextual information for large objects, or difficulties in accurately localizing objects with diverse orientations. The paper aims to bridge this gap by proposing a module that jointly addresses scale and orientation awareness for more robust and accurate detection.

Methods

The core of the proposed solution is the SOAM Block, which builds upon the AMSNet architecture. Key components include:

  • Adaptive Multi-Scale (AMS) Module: This module uses a parallel arrangement of depth-wise convolutional kernels with varying sizes (3x3, 5x5, 7x7, 9x9) to extract multi-scale features without dilation. This approach preserves local context and enhances scale adaptability. Features are fused using 1x1 convolutions and a bidirectional fusion strategy.
  • Strip-based Context Attention (SCA) Module: This module enhances long-range dependency modeling by combining global average pooling with horizontal and vertical 1D strip convolutions. This helps capture global semantic relationships and improves contextual understanding, especially in cluttered scenes.
  • Alignment Convolution (AlignConv): To address geometric misalignment between sampling grids and object orientations, AlignConv dynamically adjusts sampling positions based on anchor box geometry. This ensures better alignment of receptive fields with rotated objects.
  • Feature Alignment Module (FAM): This module integrates an anchor refinement subnetwork with AlignConv to refine spatial alignment between features and object regions, particularly for targets with orientation and scale variations.

The overall architecture combines these modules to achieve a synergistic effect, improving both local feature extraction and global context modeling.

Datasets

The authors evaluated their method on three widely recognized benchmark datasets for remote sensing object detection:

  • DOTA: A large-scale dataset with 2806 aerial images and over 188,000 object instances across 15 categories, known for its scale variation, diverse orientations, and dense targets.
  • HRSC2016: A dataset specifically for ship detection, featuring elongated targets with significant orientation variations.
  • UCAS-AOD: A smaller dataset focusing on cars and planes, designed to evaluate detection under sparse and dense conditions with arbitrary object orientations.

Key Insights and Ideas

  • Scale-Orientation Coupling is Crucial: The paper's central insight is that object scale and orientation are intrinsically linked in remote sensing imagery, and modeling this coupling is key to improving detection accuracy, especially for large objects.
  • Parallel Multi-Scale Kernels Outperform Dilated Convolutions: Using parallel depth-wise convolutions with varying kernel sizes (AMS module) is shown to be more effective than dilated convolutions for capturing multi-scale features without introducing sparsity or losing fine-grained details.
  • Strip Convolutions for Long-Range Context: The SCA module effectively leverages 1D strip convolutions to capture long-range spatial dependencies, which is vital for understanding context in complex aerial scenes.
  • Anchor-Guided Alignment is More Robust: AlignConv, guided by anchor box geometry, provides a more stable and interpretable way to align features with rotated objects compared to implicit offset learning methods.

Key Results

The proposed SOAM Block, integrated into the AMSNet architecture, demonstrated state-of-the-art performance across the evaluated datasets:

  • DOTA: Achieved a mean Average Precision (mAP) of 81.72%, outperforming existing methods like LSKNet-T (81.37%) and RVSA-ORCN (81.01%). It showed significant gains in specific categories like Plane and Bridge.
  • HRSC2016: Achieved a mAP of 92.43%, significantly outperforming models like ReDet (90.46%) and S2ANet (90.17%).
  • UCAS-AOD: Achieved a mAP of 91.86%, outperforming methods like DAL (89.87%) and SLA (89.44%).

The method also showed a favorable trade-off between accuracy and computational efficiency, with a compact model size and reduced parameter count compared to some high-performing backbones.

Limitations and Future Work

The authors acknowledge several limitations:

  • Performance Degradation: Performance may degrade in cases of extreme object size imbalance or severe occlusion, where feature misalignment can still occur.
  • Manual Kernel Size Selection: The kernel sizes in the AMS module are manually selected and might not generalize optimally across all datasets.
  • Optical Imagery Focus: The current evaluation is limited to optical remote sensing images.

Future work directions include:

  • Exploring the generalizability of the method to other modalities like SAR or multispectral imagery.
  • Investigating automatic methods for kernel size selection.
  • Developing strategies to further improve robustness against extreme size variations and occlusion.

Summary updated

BibTeX
@article{chen2025soam,
  author = {Yi Chen and Zhidong Wang and Zhipeng Xiong and Yufeng Zhang and Xinqi Xu},
  title = {{SOAM Block: A Scale-Orientation-Aware Module for Efficient Object Detection in Remote Sensing Imagery}},
  journal = {Symmetry},
  year = {2025},
  doi = {10.3390/sym17081251},
}
Figure