Reproduce “OV-DEIM: Real-time DETR-Style Open-Vocabulary Object Detection with GridSynthetic Augmentation” 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

Leilei Wang, Longfei Liu, Xi Shen, Xuanlong Yu, Ying Tiffany He, Fei Richard Yu, Yingyi Chen

Researchers developed OV-DEIM, a real-time DETR-style open-vocabulary object detector that achieves high detection accuracy and efficiency. This framework improves recognition of rare and long-tail categories through a novel GridSynthetic augmentation and a query supplement strategy, outperforming YOLO-style models in speed and performance on LVIS and COCO datasets.

Key insights

  • DETR-style architectures can achieve real-time open-vocabulary object detection performance and efficiency competitive with, or superior to, YOLO-style models, while maintaining an NMS-free, end-to-end inference process.
  • Explicitly generating additional object candidates from encoder features (query supplement) can substantially improve Fixed AP metrics without computational overhead, addressing a common limitation of fixed-query DETR models.
  • Structured data augmentation like GridSynthetic, by creating scenes with clear object boundaries and diverse co-occurrence patterns, effectively enhances vision-text alignment and semantic discrimination, leading to improved performance on long-tail categories.

Summary by alphaXiv.

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