Fig. 2: The main architecture of our proposed RiO-DETR. The framework highlights
three key components: (A) Content-Driven Angle Estimation, (B) Rotation-Rectified
Orthogonal Attention, and (C) Decoupled Periodic Refinement.