BibTeX entry

@PHDTHESIS{199912Gareth_J_Edwards,
  AUTHOR={Gareth J Edwards},
  TITLE={Learning to Identify Faces in Images and Video Sequences},
  SCHOOL={University of Manchester},
  YEAR=1999,
  URL={http://www.bmva.org/theses/1999/1999-edwards.pdf},
}

Abstract

We present novel methods for locating, tracking and interpreting faces in images and video sequences using 2D Appearance Models of faces. We describe how to construct models that represent both the shape and texture variation in faces and can be used to generate photo-realistic synthetic reconstructions of new faces. We describe how Appearance Models can be combined with an active search algorithm. We show how these Active Appearance Models (AAMs) can be efficiently fitted to image data. AAMs provide the basis for many types of analysis, including face identification and expression recognition. We show how Appearance Models can be partitioned into sub-spaces describing different types of ‘real-world’ variation such as identity, pose, lighting and expression. This partitioning provides the basis for an adaptive tracking scheme that exploits the fact that identity must remain constant during a sequence. The scheme provides online refinement of the sub-spaces during tracking and improves the stability of measurements of identity. All the methods have been systematically tested using still images and video sequences. We show that AAMs provide an effective means of interpreting faces in images and video and that the adaptive tracking scheme results in improved face recognition compared with equivalent static methods.