RiO-DETR: First End-to-End Real-Time DETR for Oriented Object Detection
jiqizhixin · x · 2026-08-23
University of Science and Technology of China, Huawei, and Tsinghua University present RiO-DETR, accepted as an Oral at ECCV 2026. It is the first end-to-end real-time DETR for oriented object detection.
Key Innovations:
- Models rotation angle as a periodic variable determined by visual content rather than just another box coordinate.
- Redesigns matching and attention mechanisms to avoid computational blowup, preserving DETR's NMS-free, end-to-end training pipeline.
Performance:
- Speed: 2.7ms (4M param variant) to 29.9ms on DOTA-1.0.
- Accuracy: Achieves 78.4 AP50 (small model) and 81.8 AP50 (large model), outperforming all prior real-time oriented detectors.
More from Research
- TwIL-LM2 1.7B outperforms larger models in strict formal reasoning — HopefulMode429 · 2026-08-23
- Magnetic resonance force microscopy enables single-nm resolution imaging — zetalyrae · 2026-08-23
- Osaka Metro Univ AI replicates captain steering, follows unlearned rules — 4310sy · 2026-08-23
- SWE-bench Science Released to Benchmark Coding Agents on Scientific Software Engineering Tasks — geoffwolfe · 2026-08-23
- Tsinghua Researchers Use GPT to Solve 40-Year-Old Optimization Problem — 新智元 · 2026-08-23
- Humanoid robotics shift toward end-to-end control for direct perception-driven movement — aftahi_ai · 2026-08-23