Reproduce “Semi-DETR: Semi-Supervised Object Detection with Detection Transformers” 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
Jiacheng Zhang, Xiangru Lin, Wei Zhang, Kuo Wang, Xiao Tan, Junyu Han, Errui Ding, Jingdong Wang and 1 more
Semi-DETR is a comprehensive framework that integrates Detection Transformers (DETR) into semi-supervised object detection, addressing the issues of noisy pseudo-labels and query consistency. The method incorporates stage-wise hybrid matching, cross-view query consistency, and cost-based pseudo-label mining, achieving 30.5 mAP on MS COCO with only 1% labeled data.
Key insights
- DETR's one-to-one bipartite matching is inefficient and less effective when applied to noisy pseudo-labels in semi-supervised settings, indicating a need for more flexible assignment strategies.
- The inherent lack of deterministic correspondence for object queries in DETR architectures poses a challenge for applying standard consistency regularization across different augmented views.
- Dynamically identifying and leveraging reliable pseudo-labels based on matching cost can significantly improve the quality and stability of training in semi-supervised object detection.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.