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 rotated usually use ln-norm losses (L1 loss, L2 smooth L1 loss) loss functions. As mainly based on non-scale-invariant Minkowski distance, using will lead to inconsistency with the metric rotational Intersection-over-Union (IoU) training instability. To address problems, we Mahalanobis distance calculate between predicted target vertices' vectors, proposing new function called Distance Loss (MDL) for detection. scale-invariant, MDL more consistent stable during than losses. alleviate problem boundary discontinuity like all other methods, further take minimum value make continuous at cases. We achieve state-of-art performance DOTA-v1.0 proposed method MDL. Furthermore, compared experiment that uses find performs better