Reproduce “Semi-Supervised Facial Expression Recognition based on Dynamic Threshold and Negative Learning” 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
Zhongpeng Cai, Jun Yu, Wei Xu, Tianyu Liu, Jianqing Sun, Jiaen Liang
This work introduces a semi-supervised learning framework for facial expression recognition, integrating a dynamic threshold adjustment for pseudo-labeling and a selective negative learning module to utilize low-confidence unlabeled samples. The approach achieved state-of-the-art accuracy on the RAF-DB and AffectNet datasets, notably surpassing fully supervised methods with significantly less labeled data, such as a 2.2% improvement over previous benchmarks on AffectNet with 2000 labels.
Key insights
- Effectively leveraging unlabeled data through dynamic pseudo-labeling and negative learning allows semi-supervised FER models to outperform fully supervised counterparts with substantially fewer labeled examples.
- Category-specific dynamic thresholds are essential for robust pseudo-labeling in FER due to varying recognition difficulties and inherent confidence levels across different emotion categories.
- Low-confidence unlabeled samples, typically discarded, can provide valuable 'complementary label' information that, when utilized via negative learning, helps refine model decision boundaries and improve generalization.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.