Rotated Object Detection via Scale-Invariant Mahalanobis Distance in Aerial Images

Rotated object detection in aerial images is a meaningful yet challenging task as objects are densely arranged and have arbitrary orientations. The eight-parameter (coordinates of box vectors) methods in rotated object detection usually use $l_{n}$ -norm losses (L1 loss, L2 loss, and smooth L1 loss) as loss functions. As $l_{n}$ -norm losses are mainly based on nonscale-invariant Minkowski distance, using $l_{n}$ -norm losses will lead to inconsistency with the detection metric rotational Intersection-over-Union (IoU) and training instability. To address the problems, we use Mahalanobis distance to calculate loss between the predicted and the target box vertices’ vectors, proposing a new loss function called Mahalanobis distance loss (MDL) for eight-parameter rotated object detection. As Mahalanobis distance is scale-invariant, MDL is more consistent with detection metric and more stable during training than $l_{n}$ -norm losses. To alleviate the problem of boundary discontinuity like all other eight-parameter methods, we further take the minimum loss value to make MDL continuous at boundary cases. We achieve state-of-the-art performance on DOTA-v1.0 with the proposed method MDL. Furthermore, compared to the experiments using smooth L1 loss and approximate SkewIoU loss, we find that MDL performs better in rotated object detection.

I. Introduction

Object detection in remote sensing has a wide range of applications in city planning, disaster rescue, and military field, which has been developing rapidly nowadays [1], [2], [3]. Different from detecting horizontal objects in general images, object detection in aerial images focuses more on objects’ orientations as objects are densely arranged and have arbitrary orientations. Therefore, rotated object detection now is applied to aerial images for high-precision detection.