Search papers, labs, and topics across Lattice.
RiO-DETR addresses the challenges of adapting DETR to oriented object detection (OBB) by introducing task-native designs for real-time efficiency. It uses Content-Driven Angle Estimation and Rotation-Rectified Orthogonal Attention for reliable orientation, Decoupled Periodic Refinement with Shortest-Path Periodic Loss for stable learning, and Oriented Dense O2O for faster angle convergence. Experiments on DOTA-1.0, DIOR-R, and FAIR-1M-2.0 demonstrate RiO-DETR achieves a new speed-accuracy trade-off for real-time oriented detection.
RiO-DETR makes real-time oriented object detection with transformers a reality by cleverly decoupling angle estimation and injecting angular diversity into dense supervision.
We present RiO-DETR: DETR for Real-time Oriented Object Detection, the first real-time oriented detection transformer to the best of our knowledge. Adapting DETR to oriented bounding boxes (OBBs) poses three challenges: semantics-dependent orientation, angle periodicity that breaks standard Euclidean refinement, and an enlarged search space that slows convergence. RiO-DETR resolves these issues with task-native designs while preserving real-time efficiency. First, we propose Content-Driven Angle Estimation by decoupling angle from positional queries, together with Rotation-Rectified Orthogonal Attention to capture complementary cues for reliable orientation. Second, Decoupled Periodic Refinement combines bounded coarse-to-fine updates with a Shortest-Path Periodic Loss for stable learning across angular seams. Third, Oriented Dense O2O injects angular diversity into dense supervision to speed up angle convergence at no extra cost. Extensive experiments on DOTA-1.0, DIOR-R, and FAIR-1M-2.0 demonstrate RiO-DETR establishes a new speed--accuracy trade-off for real-time oriented detection. Code will be made publicly available.
