The midway face is an averaging of both the shape and color of two face images. To do this, we first collect corresponding points--eye to eye, nose to nose--on each image, making sure that each point is annotated in the same order in both photos. With the two sets of points, we find the Delaunay triangluation of the correspondences' midpoints. For each triangle in the triangulation, we compute the affine transformation matrix that gives us the (x, y) coordinates for each pair of points (x', y') in the midway triangle shape, and copy over the pixel at (x, y) from the source image.
How similar do my sisters and I look? Only midway faces will tell.
The morph sequence builds upon the same idea as the midway face, only we now create 46 frames of midway faces with increasing weights on the shape morph and the color blend in each subsequent frame. Both the shape and color paramters lie between [0, 1]; in the starting frame, they are both equal 0, and in the ending frame, they are both equal 1. Once we tie these frames together in a timed sequence, we get a blend between one face to the other, and all the intermediate frames in bewteen.
What does a "mean face" of a population look like? Here, we average 100 faces from this annotated dataset of faces to create a "mean face". We also show some single faces morphed into the average population shape.
Here, we've morphed sequences between a single face and the population average.
Below, we show warped faces using the averaged photo of the population and a photo of myself.
Besides creating midway faces and morphs, we can create wonky caricatures by bringing one image further away from the other. Here, we bring my face further away from the "mean face" by adding the weighted difference between the population shape and my own face shape to the population shape before computing the warp.
The following images are of the average Taiwanese male and me, morphing into each other. First, we try warping just the shape, then just the color, and finally with both the shape and color.