PRA-Det: Anchor-Free Oriented Object Detection With Polar Radius Representation

Oriented object detection typically adds an additional rotation angle to the regressed horizontal bounding box (HBB) for representing the oriented bounding box (OBB). However, existing oriented object detectors based on regression angles face inconsistency between metric and loss, boundary discontinuity or square-like problems. To solve the above problems, we propose an anchor-free oriented object detector named PRA-Det, which assigns the center region of the object to regress OBBs represented by the polar radius vectors. Specifically, the proposed PRA-Det introduces a diamond-shaped positive region of category-wise attention factor to assign positive sample points to regress polar radius vectors. PRA-Det regresses the polar radius vector of the edges from the assigned sample points as the regression target and suppresses the predicted low-quality polar radius vectors through the category-wise attention factor. The OBBs defined for different protocols are uniformly encoded by the polar radius encoding module into regression targets represented by polar radius vectors. Therefore, the regression target represented by the polar radius vector does not have angle parameters during training, thus solving the angle-sensitive boundary discontinuity and square-like problems. To optimize the predicted polar radius vector, we design a spatial geometry loss to improve the detection accuracy. Furthermore, in the inference stage, the center offset score of the polar radius vector is combined with the classification score as the confidence to alleviate the inconsistency between classification and regression. The extensive experiments on public benchmarks demonstrate that the PRA-Det is highly competitive with state-of-the-art oriented object detectors and outperforms other comparison methods.

I. Introduction

Object detection is a fundamental visual analysis task aimed at recognizing and locating objects [1]. Early research mainly focused on horizontal object detection in natural scenes. However, instances in aerial images are densely distributed in any direction. Horizontal object detection causes misalignment between the HBBs and the objects in any orientation. HBBs introduce massive irrelevant information in the area of objects distributed in any direction, which will seriously interfere with the good performance of object detection. In addition, the HBBs cannot accurately and uniquely locate the object in any orientation. Therefore, oriented object detection gradually emerged. In recent years, oriented object detection has drawn increasing attention due to the demands of different scenes including aerial images [2], [3], scene texts [4], [5], etc [6], [7]. Existing oriented object detectors usually follow the general object detector paradigm and represent OBBs by adding additional rotation angles to the HBBs. Consequently, oriented object detection methods widely use the five-parameters to represent OBBs. Despite the satisfactory results, oriented object detection based on regression angles often faces some new issues. Unlike horizontal detection, the Intersection over Union (IoU) of two OBBs is non-differentiable for learning [8]. Due to the PoA and the EoE, oriented detectors based on angle regression often suffer from boundary discontinuity and square-like problems [9]. PoA and EoE can cause suboptimal predictions outside the defined range. Due to the sharp increase in loss at the boundary, inconsistent regression at the boundary and non-boundary during training can easily lead to training instability. Therefore, these issues greatly affect the performance of the oriented detectors.