Abstract

Locating semantically meaningful parts in visual objects (e.g., face, body) constitutes the first and fundamental step towards the automatic analysis of both the object properties and behaviours. Statistical Component analysis comprises a set of machine learning and algebraic techniques that find components under particular constraints with respect to the problem at-hand (e.g., clustering, predictive analysis and spatio-temporal alignment). These are the main axes around which my research revolves. In this talk, I present a collection of novel component analysis techniques aimed towards solving a variety of problems arising in fields such as automatic face analysis (including face and facial expression and affect recognition), surveillance, 3D object reconstruction and beyond. I also argue that the first and fundamental step for all the above applications is deformable object alignment (e.g., aligning a model of semantically meaningful parts, such as eyes, mouth etc. in case the object is human face). I will summarize my team’s contribution in deformable object alignment providing in the same time a short tutorial on some of the basic ideas in the field. Finally, I will conclude my talk by summarizing some of the open problems and challenges (such as tackling the bottleneck of relying on human annotations).