Reproduce “Dual-R-DETR: Resolving Query Competition with Pairwise Routing in Transformer Decoders” 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
Ye Zhang, Qi Chen, Wenyou Huang, Rui Liu, Zhengjian Kang
Researchers from the University of Pittsburgh and collaborators developed Dual-R-DETR, a framework that integrates pairwise routing into Transformer decoders to resolve query competition in DETR-style object detectors. This approach improves object detection mAP by up to 1.7% on COCO with ResNet-50 and 0.8% with Swin-T, achieving competitive performance without additional inference overhead.
Key insights
- Explicitly modulating pairwise query interactions during training, differentiating between inhibitory (suppressor) and facilitative (delegator) roles, significantly enhances query specialization and coverage.
- Asymmetric gating, based on dynamic query characteristics like confidence and spatial overlap, allows for more intelligent and adaptive regulation of query-to-query competition and cooperation.
- Integrating a specialized training-time mechanism that can be removed for inference enables advanced query dynamics without compromising the real-time performance of DETR models.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.