Image References: https://www.pinterest.com/pin/167266573647237114/. https://www.google.com/url?sa=i&url=https%3A%2F%2Fcommons.wikimedia.org%2Fwiki%2FFile%3AJoe_Biden%2C_official_photo_portrait%2C_113th_Congress.jpg&psig=AOvVaw0UC2HqGZJ4vp9FPRdUZc_A&ust=1583383685532000&source=images&cd=vfe&ved=0CAIQjRxqFwoTCNjYhIWCgOgCFQAAAAAdAAAAABAD
The Mid-Way Face: Below is a picture of Daenery Targaryen and Joe Biden, as well as the mid-way point between the two images:
Face Morphing: Here is the morphing GIF between the two images:
Morphing to the Mean: I used Danes. Below are several pairs of faces whose key points were morphed into the mean key points over the dataset (the first image is the original, the second morphed):
Mean Face: Below is the face that is obtained by first morphing each image into the mean key point alignment. Then, we take the pixel-wise mean over the aligned images to obtain a "mean" face:
Mean Face: Below is the face that is obtained by first morphing each image into the mean key point alignment. Then, we take the pixel-wise mean over the aligned images to obtain a "mean" face:
Caricatures: Now, we can morph Dany into the mean key point configuration, and we can morph the mean face into Dany's key point config. Note that because the two images are of different sizes, I only morph the facial regions and leave the outside black:
Bells and Whistles: Deep Random Bijective Interpolation I explored the usage of a different interpolation/ morphing algorithm. Specifically, I first instantiate a sequence of random affine transformation that operates on the full image. Then, I compute the inverse of each transform. This essentially amounts to a deep random neural network and its inverse. I then embed each image in the random feature space via a forward pass in the network. Then, in the feature space I perform linear interpolation before decoding back to an image using the inverted forward pass. Ultimately, I found that there were lots of numerical instabilities when making the network too deep, so the results below come from a shallow, linear network. Below are the results when using this method on the Dany and Biden images (the resolution is low to ensure that the inverse can be computed reasonably quickly):
