Reproduce “Exploring Facial Expression Recognition through Semi-Supervised Pretraining and Temporal Modeling” 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

Jun Yu, Zhihong Wei, Zhongpeng Cai, Gongpeng Zhao, Zerui Zhang, Yongqi Wang, Guochen Xie, Jichao Zhu and 1 more

Researchers from the University of Science and Technology of China developed a two-phase approach for facial expression recognition, integrating semi-supervised pretraining with a debiased feedback strategy and a Transformer-based temporal encoder. Their method achieved an average F1 score of 44.43% on the 6th Affective Behavior Analysis in-the-Wild (ABAW) competition validation set, demonstrating improved performance by effectively addressing data scarcity, class imbalance, and the dynamic nature of expressions.

Key insights

  • Leveraging large-scale unlabeled face recognition data through semi-supervised learning significantly boosts facial expression recognition performance, achieving a 16.96 F1 score improvement over baselines.
  • Temporal modeling with a Transformer-based encoder effectively captures dynamic expression changes, adding 2.81 F1 points, while post-processing further enhances temporal consistency.
  • A debiased feedback learning strategy is crucial for semi-supervised learning, ensuring unbiased feature learning from disparate datasets (FR vs. FER) by refining pseudo-label generation.

Summary by alphaXiv.

Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.