Database for Learning Appearance Face Models

The project started with carrying out data acquisition experiments on live video streams and preliminary evaluation on the data at the end users' site. This has helped us in identifying realistic imaging conditions and corresponding constraints under which a machine vision system can operate outside the usual laboratory environment. For learning a face model for real-time detection and tracking, publically available face databases were utilised [9, 10]. These were from Olivetti Research Lab. UK; the USENIX Faces Archive, USA; the Universities of Essex and Manchester, UK; Bern and Bochum, Germany; MIT, CMU, Harvard, USA; and Weizmann Institute, Israel. The following parameters were studied:

1. the ratio of the size of a face to the overall video-frame size, i.e. the accessible camera field of view with respect to the acceptable resolution for recognition,
2. the acceptable level of lighting variations under which the face detection, tracking and normalisation processes of the system can be performed consistently for both training and recognition tasks,
3. the required size of databases for moving faces to be used for learning temporal models for recognition in dynamic scenes,
4. determining reference feature points on all faces to be used for bootstrapping view alignment.

Further studies were carried out on identifying methods and sources of variation in the acquisition of face database for both learning a view-based face appearance model and applying such a model for face recognition across views [5, 7, 14]. More recently, a novel approach for automatic acquisition of a face database across views has been developed. Such a database has been used for learning generic models for real-time head pose estimation and tracking [21], and facial identity recognition and tracking across views [8].