Arjun's CS294-26 Proj3

Part 1 Arjun <-> Steph Face Morph

I labeled 48 correspondences betweeen my face and Steph Curry's. Then I computed the mean location for each correspondence and computed the Delaunay triangulation from there. The results are shown below.


Below we show the result of warping each image to the mean correspondence location then taking an even cross dissolve. Middle is the morph, left and right the originals.


And here's 1.5s 30fps GIF morphing my face to Steph's, using 45 intermediate morphs.

Part 2 Danes Mean Person

To compute the mean person, we warp each color face to the mean keypoints. Then cross dissolve weighing each image equally.


And below we have three random people in the dataset with their faces warped to the mean keypoints.


We also computed the mean man and woman. These will come in hand later in Part 3.

Part 3 Morphing Me with Danes

Left is me warped to the keypoints of the Danes mean. Right is the Danes mean mapped to my keypoints. Forehead is a bit strange, since the dataset doesn't have any keypoints above the eyes. Eyes to chin however seems convincing.


Next are caricatures of me (both color and shape). Left to right is alpha=0.4, 0.7, 1.0. Alpha 0.4 is pretty nice, with 1.0 similar to oversharpening we start to see some unnatural distortion.


Finally we make me look more female. Bell's and Whistles The strategy is to morph my image with the male and female means. Weighing myself 1, the female mean alpha, and the male mean -alpha. Again left to right is alpha=0.4, 0.7, 1.0. Similarily Alpha 0.4 seems nice with 1.0 seeming unnatural again.