Generative and discriminative face modelling for detection

This paper reports a new image model combining self mutual information based generative modelling and fisher discriminant based discriminative modelling. Past work on face modelling have focused heavily on either generative modelling or boundary modelling considering negative examples. The motivation of this work is to examine the combinational treatment and study its effect. To effectively learn the model's parameters, a tree structure is employed to describe the inter-pixel relationships, both due to the simplicity of the structure representation and the ease of parameter estimation through decoupling a full distribution into pair-wise distributions. To fit training data distribution more accurately, we use a non-parametric representation rather than a particular parametric family of distributions for entropy estimation. We explicate the model learning and demonstrate its effectiveness primarily through the problem of face detection, i.e. modelling the 2d image appearance of human face.