Some Fun Face Research

Hi there! This page began as a COMP 254 assignment and just kept growing. We got to pick our own final project, and I chose to do a landmark-based statistical shape study of faces.

• Introduction and Background
• The Data -- It can be hard to find images of faces under controlled lighting and with uniform size/resolution. I experimented with several face databases and rejected all but one of them eventually.
• UNC has a large set of face images (about 360 men, 70 women). Unfortunately they are small and privacy issues restrict their use. They also lack controlled pose and lighting. Eventually I may do some work this database, but I didn't use it for this project.
• The Face Recognition Home Page has pointers to several face databases. Unfortunately, most of the databases were unsuitable for one reason or another. The USENIX xfaces database, for example, has images that are too small.
• The Carnegie Mellon Computer Vision Page has links to many test databases. Carnegie Mellon has a huge set of pictures of people, but the faces occur without any controlled pose.
• Olivetti Research Laboratory (ORL) has a good database (40 people) in different poses. The images are small (92x112), but the lighting is controlled and each person has several poses. One problem is that only four women occur in the entire database, so I had little choice in selecting images of women.
• Tools -- My personal feeling is that you can never have too many good tools. I found a number of packages and programs useful for exploring this idea. Here are a few of the things that I used.
• IGLOO, an object-oriented library for image processing. IGLOO is the brainchild of Dr. Coggins (my instructor) here at UNC.
• edgewarp from Bookstein's project. The source code is available with minimal copyright conditions. Some screen shots of the menu and the interface are here.
• The KUIM Image Processing System, written by John Gauch, is based on a package called /usr/image, also from UNC. Gauch was a student here before he went to Kansas University.
• xv, an image manipulation tool from John Bradley. This is possibly the handiest "swiss army knife" of handling pictures.
• The Procrustes Metric -- Once feature points are created for each face, they must be aligned. The alignment involves finding the centroid of the features, translating to the origin, normalizing the scale of the face, and then applying the procrustes metric. The "procrustes metric" refers to rotating a set of points to match another point set such that a distance metric is minimized.
• It is interesting to note the origin of the name of this metric. The name is taken from Greek mythology. Procrustes was a highway robber and son of Poseidon who forced travelers to fit into his bed by stretching their bodies or cutting off their legs until they fit into the bed. He was killed by Theseus, who gave him the same punishment he had dealt out to his victims.
• Methods/Procedures
• Results
• Lessons, or things I would do differently.
• Use bigger (color) images. If the input images are small and grayscale, you've already reduced the impact and insight that the final results can give.
• Data can be hard to get, and harder to make. Dr. Coggins says that data is harder to get BECAUSE its so hard to make, which makes sense if you think about it.
• Blending images is better than landmarks. People seem to have perceptual "circuits" that respond much more to images than to vector graphs of feature points.
• One person should do all the feature point selection to ensure more consistent data.
• Beware the file format swamp--it can suck you in. I had to handle the following formats during my work: gif, tiff, ppm, pgm, IGLOO im, KUIM im, rgb, and PostScript. A good suite of image utilities is essential.
• Face stuff is cool! I loved this project and plan to explore it more when I can find better data.
• Future Work
• With some touch-up work, this technique could generate some real research or an interesting paper. The most pressing need would be for better data. The researchers at St. Andrews University have published papers on their database of 162 faces. With my graphical interface, I can do a face in 15 minutes. With some more work on the GUI, I could get that down to 5-10 minutes. If I had high-quality face images, it would take roughly a workweek (40 hours) to get 175-200 faces digitized.
• Artificial aging/youthing by shape and color.
• Using bending energy plus an image similarity metric to "guess" a person's age.
• Beautfication, hyper-beautification (i.e. caricature of beauty), androgynous-nization, and gender changes.
• Gradual interpolation of landmarks + Thin Plate Splines = morph movies of all of the above. You can "gradually" age a person.
• Automatic feature classification. This would be quite difficult.
• References

Matt Cutts