Reproduce “DETRs with Hybrid Matching” 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
Ding Jia, Yuhui Yuan, Haodi He, Xiaopei Wu, Haojun Yu, Weihong Lin, Lei Sun, Chao Zhang and 1 more
DETRs with Hybrid Matching proposes a training scheme that combines one-to-one and one-to-many label assignment to address the sparsity of positive training samples in DETR-based models. The approach consistently improves performance across 2D/3D object detection, pose estimation, and tracking tasks, notably boosting Deformable-DETR (Swin-L) to 59.4% mAP on COCO val.
Key insights
- Increasing the number of positive training samples through a hybrid matching scheme significantly improves the accuracy and robustness of DETR-based models.
- The proposed method maintains DETR's NMS-free, end-to-end inference as the one-to-many matching branch is only active during training, ensuring no inference overhead.
- Improved optimization of the Transformer encoder, facilitated by the richer supervision from hybrid matching, is crucial for the observed performance gains.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.