The objective of this research is to understand the structure of face space, defined as the set of all images of faces under different viewing conditions. The motivation of modelling and exploring the face space is to do robust face recognition, which is still a difficult problem after decades of work. Also we try to investigate some phenomena more qualitatively, e.g. two people look more alike than images of same person under different viewing conditions.

The key idea of the work is: synthesize images under different viewing conditions, then use different techniques to model and explore the space of face images. In the preliminary work, we show how face space may be modelled and explored; we show that distance is a good measure of the separability of two classes; we also determine the viewing conditions that are best (or worst) for face recognition.

Our future work includes: investigate other observations, add new variabilities to face space by image rendering and so on.

By ambiguous we mean the condition where it is hard to perform face recognition. This is achieved by computing the intersection of the tangent planes in the face space.

Figures 1 and 2 show the ambiguous pose and lighitng under which it is hard to perform recognition.

Visualize the curvature of the face space by plotting the normalized curvature in the parameter space, which is called the Curvature Map in the thesis. Note that the high curvature means that face images change dramatically when the viewing condition changes.

Figures 3 and 4 show some face images with high and low curvature under varying pose and illumiantion.