FACE MORPHING

FACE MORPHING

Computational Photography

by Emine Begum Gulsoy

This project involves morphing two similar images into each other thorough defining corresponding points and creating a matching triangulation from these points and finally inverse warping one image into the other. Photos of class members were taken and their corresponding points were defined using MATLAB and IDL. We were supposed to use 60 frames in morphing our own pictures to the next person in line. However the program can easily be altered, through an input parameter, to give any number of frames. The morphed image series were put together as an animated gif using Adobe Photoshop. Delaunay triangulation was preferred while creating triangles through the points defined. The following in the results and discussion on how these results were achieved:

THE MORPH

Left: My image

Right: Middle Image

Final Image

Animated

The hardest part of this process is defining the corresponding points. The points for the above images have been separately defined using IDL man. The points for the rest of the images under this project have been taken from the accumulated class /faces index. One other obstacle in programming- a lot less minor- has been the conversions between using IDL and MATLAB. The two languages have the opposite ways of reading in images such that the points defined by MATLAB had to be converted. In general the program works as following:

The two images, their corresponding point indices, the number of points used and the number of output frames aimed, are given to the program as the inputs. The program gets these and creates a triangulation using dalaunay approach. Using an affine transformation matrix, produced by the ‘affine’ function, applies an inverse warp to create the final image of each step. The frames are iterated by 1/(n_frames/2). Finally the images are cross-dissolved to get the correct color.

THE MEAN FACE

mean face of 13 people

Left: Mean of 5 girls in the class

Right: Mean of 8 guys in the class

The mean face was calculated using 13 people and their corresponding txt files from the faces folder that was created for the class. Then the males and females were separated to get the mean male and female faces. Since mean face averages a big number of faces, it looks a lot smoother, perfect and symmetric compared to other pictures. However we all look funny when warped into the mean shape, here it is:

EIGEN FACES

Here are the eigen faces computed using the built-in LA_svd function in IDL. The first one represents the mean face. The below images are the outputs from only the red channel. However, the program can also generate all the eigen faces of all the channels and put them together.