A Comparative Study of Loss Functions for Arbitrary-Oriented Object Detection in Aerial Images

Detecting objects with specific orientations in aerial imagery is a complex task due to the random rotations and high density of objects within the images. Additionally, boundary discontinuity from regression parametrization poses a major challenge in designing loss functions for rotated object detection. Several bounding box representations and their associated loss functions are being gradually developed to address this challenge. In this study, we compare loss functions for oriented bounding boxes including Probabilistic IoU (ProbIoU), KFIoU and others. The evaluation uses DOTA-PS, a subset of a large scale aerial benchmark dataset DOTAv1.0, which includes two categories, plane and ship. We also introduce Complete Rotated IoU (CRIoU) loss which is improved over Rotated IoU (RIoU) loss. ProbIoU and CRIoU loss achieved mAP score of 93.1% on the test set while RIoU and KFIoU losses got 92.8% and 90.6% respectively. CRIoU demonstrates marginal improvement in oriented bounding box localization, supported by the brief experimental and visual analysis.