Reproduce “RiO-DETR: DETR for Real-time Oriented Object Detection” on OpenResearch

Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.

Universal app for Apple silicon and Intel. Requires macOS 11 or later.

About this paper

Zhangchi Hu, Yifan Zhao, Yansong Peng, Wenzhang Sun, Xiangchen Yin, Jie Chen, Peixi Wu, Hebei Li and 3 more

RiO-DETR introduces the first real-time oriented detection transformer, achieving a new speed-accuracy trade-off by addressing challenges like semantic-geometric coupling, angular periodicity, and slow convergence. The model demonstrates superior performance on benchmarks such as DOTA-1.0, DIOR-R, and FAIR-1M-2.0, with RiO-DETR-n achieving 78.4 AP50 at 2.7 ms latency on DOTA-1.0.

Key insights

  • Decoupling angular information from positional queries, allowing content embeddings to infer orientation, improves the robustness of angle prediction.
  • Designing iterative angle updates and loss functions that explicitly account for the periodic nature of angles is critical for stable and accurate refinement.
  • Artificially increasing angular diversity in training data through discrete rotations significantly accelerates the learning of orientation without incurring additional computational overhead.

Summary by alphaXiv.

Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.