Face Morphing

by Taylor Wong

Overview

In the first part of this assignment we produce a morph of a face or object into another face or object. In order to create a smooth transition from one face to the other, we combine two techniques-- a simultaneous warp of the image shape and a cross-dissolve of the image colors. The warp is performed by applying an affine transformation to the correspondences between the two images.

Image 1	Image 2

Defining Correspondences

We start by defining pairs of corresponding points between the two images. Then we compute a Delaunay triangulation on the midway of the two point sets (to create a triangulation that works equally well for both of the sets, reducing the number of deformed triangles).

Image 1	Midway	Image 2

Computing the Mid-way Face

Before producing the entire morph sequence, let us first compute the mid-way face, which will be the image at the halfway point of our morph sequence. As stated before, this involves warping both faces to the average keypoint location of the two faces and averaging the colors together.

To perform the affine warp, an affine transformation must be determined for each triangle between a face and the midway face. This affine transformation is used to map each of the three points in a triangle to its new location in the midway face. Then the inverse of the transformation is used to interpolate all the pixels within the boundary of that triangle from the source.

The Morph Sequence

Producing the morph sequence then becomes a matter of gradually changing our parameters for warp fraction and dissolve fraction, which control the relative contributions of each face to the averaged face.

Another Example: Andrew to Uncle Drew

Image 1	Image 2	Morph

Bells and Whistles: Seven Generations of the Porsche 911

Original (1963), G series, 964, 993, 996, 997, and the 991 (2011)

The "Mean Face" of a Population

I used the male faces from the IMM Face Database located at http://www2.imm.dtu.dk/~aam/datasets/datasets.html

To compute the average face of the subset of this dataset, we need to morph each of the faces into the average shape and then overlay them. Below are some examples.

Original	Morphed

Below is the average face of the males from the IMM Face Database. This face is so aesthetically pleasing because all of the irregularities have essentially been averaged out.

Now we show my face warped into the average geometry:

And finally the average face warped into my geometry:

Bells and Whistles: Morphing my gender and ethnicity

Me	Average German Woman	Shape-only morph	Appearance-only morph	Complete morph