Introduction and Background

Introduction

Is this technique new? F. Galton (see reference list) reported averaging faces as early as 1878 by taking multiple photographic exposures. To minimize the blur, he matched eye positions of each face. In retrospect, matching eyes anchors the rough position and orientation of the face, almost like a primitive procrustes normalization.

Background

• Studies of this sort might be useful for missing persons, understanding how people recognize faces, etc.

• Susan Brennan's began a study of faces, described in A. K. Dewdney's The Tinkertoy Computer. She used 186 data points, then interpolated between feature points from different people. Dewdney doesn't mention whether she warped images or only feature points. Dewdney gives an example program, FACEBENDER, to do some simple manipulations of feature points on a face.

• SIGGRAPH Technical Sketch. At SIGGRAPH 1995, Duncan Rowland gave a brief lecture about his university's work in several aspects of face landmarks. He researches at The Perception Lab, St. Andrews University, Scotland. Their web page has excellent illustrations of face landmarks, aging faces, and rating attractiveness of different faces. Rowland spoke on some things to do with face images and their landmarks.
  • Caricature -- To caricature a person, simply measure the difference between his or her landmarks and the average face, then amplify the differences. Researchers discovered that up to a limit, face recognition gets faster the more exaggerated the caricature becomes.
  • Attractiveness -- Researchers quantified the difference between average faces and attractive faces. They labeled faces as attractive or not and blended each group to produce an average face. The average "attractive" face was consistently ranked higher. An "attractive + 50%" face was ranked higher than the average and the attractive face.
  • Personal transforms -- Any measurable quality can be considered and incorporated into a statistical group. We have seen that caricature and attractiveness can be manipulated. So too can gender and age. Changing age can change shape, color, or both. With enough data, a person can pick how much to "age" a photo.