DARDet: A Dense Anchor-free Rotated Object Detector in Aerial Images

Feng Zhang,Xueying Wang,Shilin Zhou,Yingqian Wang

DOI: https://doi.org/10.1109/LGRS.2021.3122924

2021-10-03

Abstract:Rotated object detection in aerial images has received increasing attention for a wide range of applications. However, it is also a challenging task due to the huge variations of scale, rotation, aspect ratio, and densely arranged targets. Most existing methods heavily rely on a large number of pre-defined anchors with different scales, angles, and aspect ratios, and are optimized with a distance loss. Therefore, these methods are sensitive to anchor hyper-parameters and easily suffer from performance degradation caused by boundary discontinuity. To handle this problem, in this paper, we propose a dense anchor-free rotated object detector (DARDet) for rotated object detection in aerial images. Our DARDet directly predicts five parameters of rotated boxes at each foreground pixel of feature maps. We design a new alignment convolution module to extracts aligned features and introduce a PIoU loss for precise and stable regression. Our method achieves state-of-the-art performance on three commonly used aerial objects datasets (i.e., DOTA, HRSC2016, and UCAS-AOD) while keeping high efficiency. Code is available at <a class="link-external link-https" href="https://github.com/zf020114/DARDet" rel="external noopener nofollow">this https URL</a>.

Computer Vision and Pattern Recognition

What problem does this paper attempt to address?

The paper attempts to address the challenges faced in rotating object detection in aerial images. Specifically, these challenges include:

1. **Significant variations in scale, rotation, aspect ratio, and dense arrangement of objects**: Objects in aerial images often have different sizes, orientations, and aspect ratios, and are frequently densely packed together, making it difficult for traditional anchor-based methods to handle.
2. **Boundary discontinuity problem**: Existing rotating object detection methods tend to experience performance degradation when dealing with objects near boundaries, as the periodicity of angles and the interchangeability of edges can cause the loss function to increase sharply at the boundaries.
3. **Sensitivity to anchor box hyperparameters**: Most existing methods rely on a large number of predefined anchor boxes with different scales, angles, and aspect ratios. Therefore, these methods are highly sensitive to the hyperparameters of the anchor boxes, making them prone to performance degradation.

To address these issues, the paper proposes a new Dense Anchor-free Rotating Object Detector (DARDet), which directly predicts five parameters on each foreground pixel to encode the oriented bounding box (OBB). It also designs a new Aligned Convolution Module (ACM) to extract aligned features and introduces PIoU loss to optimize the regression process. Experimental results show that this method achieves state-of-the-art performance on several commonly used datasets while maintaining high efficiency.