Reproduce “Efficient Face Image Quality Assessment via Self-training and Knowledge Distillation” 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
Wei Sun, Weixia Zhang, Linhan Cao, Jun Jia, Xiangyang Zhu, Dandan Zhu, Xiongkuo Min, Guangtao Zhai
This work introduces an efficient Face Image Quality Assessment (FIQA) framework that leverages iterative self-training and multi-source knowledge distillation. The method achieves high quality prediction accuracy while drastically reducing computational complexity, securing first place in the ICCV 2025 VQualA FIQA Challenge.
Key insights
- Iterative self-training with pseudo-labels effectively enhances the teacher model's performance and generalization, even with marginal gains.
- Multi-source supervision, integrating ground-truth and pseudo-labels from different teacher iterations, is crucial for transferring comprehensive knowledge to the lightweight student model.
- The combination of self-training and knowledge distillation enables the student model to achieve performance comparable to the high-capacity teacher while drastically reducing computational requirements.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.