My research interest lie in the research areas of statistical machine learning (with emphasis on component analysis and kernel learning) with applications in computer vision and general image/signal analysis. In particular, I design and develop statistical machine learning algorithms taylored for 2D/3D object and face recognition, 2D/3D object tracking, 2D/3D deformable object/face alignment, as well as, automatic human behaviour analysis. I apply such algorithms for analysis of various other signals (e.g., brain signals and graphs).
A recent line of research in computer vision is to apply tools from statistical machine learning in order to solve challenging problems such as object recognition, object tracking, image alignment, human behaviour analysis etc. I work on this line of research on the following topics
(1) Deformable models for face alignment. Construction of deformable models provide an excellent paradigm on how elements of computer vision (i.e., 2D/3D motion models, image warping etc) can be successfully combined with elements of statistical machine learning (component analysis). My work on deformable models is on the design and application robust component analysis and optimization methodologies for building facial deformable models that can work on unconstrained conditions.
(2) Face recognition. Arguably face is the most used object in computer vision due to the numerous applications that involve faces, as well as, since it is a appears in very challenging settings (i.e., it is highly deformable, it has drastic appearance changes due to different illumination/poses/aging/cosmetics etc.). My work concentrated on the development of robust algorithms for face recognition in challenging real-world application (e.g., surveillance).
(3) Human behaviour analysis. Automatic understanding of human behaviour and especially non-verbal facial behaviour from visual input is a key aspect in the majority of modern human computer interaction systems. I am very active in this line of research working towards the design of algorithms that take into account both the spatial, as well as, the temporal (dynamic) nature of human behaviour.
Statistical Machine Learning
Component analysis, such as Principal Component Analysis, Linear Discriminant analysis, etc is among the most well-researched and popular topics of statistical machine learning. I am working on various aspects of linear/non-linear/multilinear, as well as, probabilistic and deterministic component analysis. My research in component analysis can be divided in the following paths
(1) Defining new component analysis algorithms. One of my main lines of research is the development of new component analysis for the problems at hand (e.g., incorporating a-priori knowledge etc) and theoretically studying their properties (e.g., convergence)
(2) Robust component analysis. In many cases, due to the similarity measures employed in the definition of the optimization problem which produces the components (i.e., L2 norm etc) component analysis algorithms are very sensitive to noise/outliers. One of my main lines of research is to develop robust component analysis algorithms by designing robust kernels (both positive definite and indefinite).
(3) Probabilistic component analysis is a very powerful framework that naturally allows the incorporation of noise and a-priori knowledge in the developed models. One of my recent lines of my research is to develop a unified framework for probabilistic component analysis.
(3) Complex and hyper-complex valued component analysis. The majority of component analysis have been defined on real-valued data. Very limited research has been conducted on how complex-valued data can be appropriately used in component analysis. The most recent line of my research is the study of component analysis using complex and hyper-complex-valued data.