Reproduce “MixTeacher: Mining Promising Labels with Mixed Scale Teacher for Semi-Supervised Object Detection” 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
Liang Liu, Boshen Zhang, Jiangning Zhang, Wuhao Zhang, Zhenye Gan, Guanzhong Tian, Wenbing Zhu, Yabiao Wang and 1 more
Researchers from Youtu Lab, Tencent, developed MixTeacher, a semi-supervised object detection framework that addresses scale variation by improving pseudo-label quality. It employs a Mixed Scale Teacher to generate more accurate labels by adaptively fusing multi-scale features and introduces Promising Labels Mining to recover valuable low-confidence predictions, achieving state-of-the-art performance on MS COCO with up to 1.5 mAP improvements.
Key insights
- Adaptive fusion of multi-scale features within the teacher network significantly improves the quality and recall of pseudo-labels for objects across various sizes, outperforming single-scale pseudo-labeling.
- Mining low-confidence pseudo-labels based on their 'score promotion' from mixed-scale features effectively recovers valuable training signals that would otherwise be discarded, boosting overall detection performance.
- Explicitly enhancing pseudo-label generation quality is more effective for semi-supervised object detection than relying solely on consistency regularization across different views.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.