Uncovering The Subtle Details of Faces
Thesis event information
Date and time of the thesis defence
Place of the thesis defence
L2, Linnanmaa campus
Topic of the dissertation
Uncovering The Subtle Details of Faces
Doctoral candidate
Master of Science in Technology Tuomas Varanka
Faculty and unit
University of Oulu Graduate School, Faculty of Information Technology and Electrical Engineering, Center for Machine Vision and Signal Analysis
Subject of study
Computer science and engineering
Opponent
Professor Esa Rahtu, Tampere University
Custos
Professor Guoying Zhao, University of Oulu
Machine learning approaches for generating and analyzing subtle facial details
The human face provides a variety of information, including one's identity, intention, and expression. We are experts at reading other faces---distinguishing expressions from facial movements and each other from deviations in facial features. Distinguishing expressions from facial movements and discerning one from another from deviations can make the biggest difference, from the identity to the emotion. This thesis examines these subtle facial details using both discriminative and generative approaches.
Discriminative approaches map input data into classes, particularly videos that capture micro-expressions (MEs) in this thesis. MEs are rapid, involuntary, and subtle facial movements. Due to these characteristics, they are too difficult for humans to perceive; hence, automatic methods have been developed. The complexities of MEs are analyzed in-depth, addressing two primary issues: (1) the evaluation of ME recognition methods, where data leakage and inconsistent metrics are identified as factors that misrepresent performance, and (2) the significance of video information compared to stationary images in ME analysis.
Generative approaches refer to generating new data, particularly facial images in this work. Focus is given to subtlety in identity details and facial expressions, especially with two applications on face restoration and face animation. (1) Given that humans are experts with faces, the smallest details in faces can make a difference in identity. Personalizing a face restoration model is used to retain identity-related information even from very low-quality images, where facial details may be lost. (2) A method is developed to enable precise control of facial muscle movements in animations of faces.
Discriminative approaches map input data into classes, particularly videos that capture micro-expressions (MEs) in this thesis. MEs are rapid, involuntary, and subtle facial movements. Due to these characteristics, they are too difficult for humans to perceive; hence, automatic methods have been developed. The complexities of MEs are analyzed in-depth, addressing two primary issues: (1) the evaluation of ME recognition methods, where data leakage and inconsistent metrics are identified as factors that misrepresent performance, and (2) the significance of video information compared to stationary images in ME analysis.
Generative approaches refer to generating new data, particularly facial images in this work. Focus is given to subtlety in identity details and facial expressions, especially with two applications on face restoration and face animation. (1) Given that humans are experts with faces, the smallest details in faces can make a difference in identity. Personalizing a face restoration model is used to retain identity-related information even from very low-quality images, where facial details may be lost. (2) A method is developed to enable precise control of facial muscle movements in animations of faces.
Last updated: 5.5.2025