Challenges in Learning the Appearance of Faces for Automated Image Analysis: Part I
In: PASCAL Workshop on Pattern Recognition and Machine Learning in Computer Vision, 3-5 May 2004, Grenoble.
The variability of images of the human face challenges research in machine vision since its beginning. Sources of variability not only include individual appearance but also cover external parameters such as perspective and illumination that influence the image formation process heavily.
Research on the analysis of face images currently splits into the directions of face detection and face recognition. Approaches and problem setting in this two areas are still quite different despite the common goal to compensate for the large variability of faces in images. In both areas machine learning strategies are used to learn from example faces a general model of the appearance of human faces.
In the first part of our presentation we would like to review the current state of the art in face detection and in face recognition. In a second part we will compare the different strategies used and try to describe the requirements of a general image model that could serve as basis in detection as well as in recognition research.
For face detection we will focus on methods based on the estimation of the probability distribution of large number of features computed on face examples. After selecting the subset of features which appear to be most relevant to the task, faces are detected by combining the outcome of suitably defined statistical tests. This method, which is based on positive examples only, seems to give very promising results compared to state of the art techniques based on both positive and negative examples.
In face recognition we will concentrate on methods that use an analysis by synthesis framework such as morphable models, active appearance or shape models. Currently these approaches seem the most promising methods able to account for variations in perspective and illumination.