Reproduce “Deep Learning for Micro-expression Recognition: A Survey” on OpenResearch

About this paper

Yante Li, Jinsheng Wei, Yang Liu, Janne Kauttonen, Guoying Zhao

Researchers from the University of Oulu's CMVS conducted the first comprehensive survey on deep learning methods for micro-expression recognition (MER) since 2016, establishing a novel taxonomy and systematically analyzing architectures, inputs, training strategies, and loss functions. The work consolidates the field's advancements, discusses challenges in handling subtle and fleeting facial movements with limited data, and charts future research directions, serving as a foundational reference.

Key insights

• Combining multiple input modalities, especially optical flow with appearance-based features, provides complementary information crucial for recognizing subtle micro-expressions.
• Transfer learning strategies, such as fine-tuning and knowledge distillation from larger datasets or teacher models, are vital to address the extreme data scarcity in MER.
• Specialized network architectures and modules, including attention mechanisms, Graph Convolutional Networks (GCNs), and multi-task learning, are increasingly adopted to handle the unique spatio-temporal challenges of micro-expression data.

Summary by alphaXiv.