TS⁴Net: Two-stage sample selective strategy for rotating object detection

Rotating object detection has recently attracted increasing attention in aerial photographs, remote sensing images, etc. In this paper, we propose a rotating object detector TS⁴Net, which contains anchor refinement module (ARM) and two-stage sample selective strategy (TS⁴). The ARM can convert the preset horizontal anchors into high-quality rotated anchors through two-stage refinement, which also adopts a rotated Intersection-over-Union(IoU) prediction branch to improve localization accuracy in the second stage. TS⁴ module utilizes different constrained sample selective strategies to allocate positive and negative samples, which is adaptive to the regression task in different stages. Benefiting from the ARM and TS⁴, the TS⁴Net can achieve superior performance for rotating object detection solely with one preset horizontal anchor. Considering that most rotating object detection datasets mainly focus on the field of remote sensing and are shot in high-altitude scenes. We present a low-altitude drone-based dataset, named UAV-ROD, aiming to promote the research and development in rotating object detection and UAV applications. The UAV-ROD dataset can be used in rotating object detection, vehicle orientation recognition, and object counting tasks. Extensive experiments on the UAV-ROD dataset and four datasets demonstrate that our method achieves competitive performance against most state-of-the-art methods.