Reproduce “Point-Teaching: Weakly Semi-Supervised Object Detection with Point Annotations” 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
Yongtao Ge, Qiang Zhou, Xinlong Wang, Zhibin Wang, Hao Li, Chunhua Shen
A framework called Point-Teaching enables the training of accurate object detectors by combining a small set of bounding box annotations with extensive, cost-effective point annotations. This method significantly improves performance over existing semi-supervised approaches, achieving a 9.08 AP increase on MS COCO with only 0.5% fully-labeled data compared to Unbiased Teacher.
Key insights
- Point annotations offer a superior balance of annotation efficiency and localization information compared to image-level labels, making them ideal for weakly semi-supervised learning.
- The combination of Hungarian matching, image-wise MIL, and a novel point-wise MIL loss effectively translates sparse point cues into high-quality, instance-level supervision for robust detector training.
- Dynamically augmenting training data with 'unmatched' instances via point-guided copy-paste improves model robustness and mitigates class imbalance in generated pseudo labels.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.