Reproduce “Enhancing Your Trained DETRs with Box Refinement” 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

Yiqun Chen, Qiang Chen, Peize Sun, Shoufa Chen, Jingdong Wang, Jian Cheng

The RefineBox framework introduces a lightweight, plug-and-play network to enhance the localization accuracy of pre-trained Transformer-based object detectors. This approach improves Average Precision by up to 2.8 AP on COCO and LVIS datasets by efficiently refining bounding box predictions without modifying the base detector.

Key insights

  • Quantitative error analysis confirms that localization is the primary performance bottleneck for DETR-like models, accounting for roughly 25 AP potential gain, significantly more than classification errors (3-5 AP).
  • A dedicated, small-capacity refinement module can effectively and substantially improve localization, demonstrating that complex architectural overhauls are not always necessary for performance boosts.
  • Freezing the well-trained base detector during the refinement network's training ensures a stable feature distribution, prevents performance degradation, and allows for highly efficient learning.

Summary by alphaXiv.

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