In this assignment, interest points are manually selected on the various faces. 22 interest points were used to match one face to another. These points are then matched to one another to morph the faces together. To smoothly transition from one face image to another these interest points are cross dissolved to generate interest points of an intermediate face image. The offsets of the selected interest points between the true and the intermediate face image can then be computed by simply computing the difference between the interest points. The offsets of all the other points (non-interest points on the face) are unknown. We compute these values using Poisson fill. The following figures illustrate the chosen interest points and a "mean" face.
Specifically, for each pair of face images we compute an intermediate image. The offsets from the two original images two the intermediate image are computed by doing a Poisson fill. The Poisson fill is computed separately for computing the X and the Y coordinates. Essentially, we end up performing two backward warp operations to compute the appropriate color value for each pixel in the intermediate image, one for each of the two images being combined. The resulting color values are then combined using a convex combination of the warp results.
Here is an animation of the morphed faces:
In general, the results look alright and could be significantly improved with the correct ordering of faces. It is hard to find smooth transitions between significantly different faces or faces with different orientations. The overall mean face of the class is:
