[论文解读] Learning a Rotation Invariant Detector with Rotatable Bounding Box

Lei Liu, Zongxu Pan+1 more|arXiv (Cornell University)|Nov 26, 2017

Advanced Neural Network Applications参考文献 23被引用 165

一句话总结

本文提出 DRBox，一种使用可旋转边界框（RBox）的旋转不变检测器，以在遥感图像中准确检测任意方向的对象，并在旋转鲁棒性和角度估计方面超过 SSD 与 Faster R-CNN。它在训练阶段学习对象方向，并输出多角度对象的方向角。

ABSTRACT

Detection of arbitrarily rotated objects is a challenging task due to the difficulties of locating the multi-angle objects and separating them effectively from the background. The existing methods are not robust to angle varies of the objects because of the use of traditional bounding box, which is a rotation variant structure for locating rotated objects. In this article, a new detection method is proposed which applies the newly defined rotatable bounding box (RBox). The proposed detector (DRBox) can effectively handle the situation where the orientation angles of the objects are arbitrary. The training of DRBox forces the detection networks to learn the correct orientation angle of the objects, so that the rotation invariant property can be achieved. DRBox is tested to detect vehicles, ships and airplanes on satellite images, compared with Faster R-CNN and SSD, which are chosen as the benchmark of the traditional bounding box based methods. The results shows that DRBox performs much better than traditional bounding box based methods do on the given tasks, and is more robust against rotation of input image and target objects. Besides, results show that DRBox correctly outputs the orientation angles of the objects, which is very useful for locating multi-angle objects efficiently. The code and models are available at https://github.com/liulei01/DRBox.

研究动机与目标

• Motivate rotation-invariant object detection for arbitrarily oriented targets in remote sensing imagery.
• Propose rotatable bounding boxes (RBox) to encode orientation alongside location and size.
• Develop DRBox, a CNN-based detector that predicts multi-angle RBoxes and angles.
• Train DRBox with ArIoU-based matching to enforce correct angle estimation.
• Demonstrate improved accuracy and rotation robustness over traditional BBox-based detectors.

提出的方法

• Define rotatable bounding box (RBox) with five parameters including angle.
• Use IoU and angle-aware ArIoU for matching and training guidance.
• Extend SSD loss with an angle regression term using tangent formulation for angle periodicity.
• Incorporate multi-angle prior RBoxes to search across orientations during detection.
• Apply pyramid input and 300x300 sub-images to handle large satellite images.
• Train three specialized DRBox models for ships, vehicles, and airplanes with fixed aspect ratios tailored to each category.

实验结果

研究问题

• RQ1 Can a rotatable bounding box enable rotation-invariant detection for arbitrarily oriented objects in satellite imagery?
• RQ2 Does training with ArIoU-based matching improve angle estimation and detection accuracy across rotations?
• RQ3 How does DRBox compare to BBox-based detectors (SSD, Faster R-CNN) in precision, recall, and rotation robustness on remote sensing data?
• RQ4 What is the impact of multi-angle priors and pyramid input on detection performance and speed?

主要发现

• DRBox outperforms Faster R-CNN and SSD on ship, vehicle, and airplane detection in remote sensing images across BEP, AP, and mAP metrics.
• DRBox achieves higher precision-recall performance and better robustness to rotation of both input images and objects (STD_AP and STD_AS).
• ArIoU-based matching helps assign positive samples with appropriate angle guidance, enabling learned orientation estimation.
• DRBox outputs orientation angles for detected objects, aiding localization of multi-angle targets.
• DRBox runs at 70–80 fps on GTX 1080Ti, with pyramid input adding at most 4/3 time cost, yielding about 1600x1600 px2/s.