CS 180 Project 3 - Face Morphing and Modelling a Photo Collection by Eshani Jha

In this project, I explored face morphing techniques, including warping facial features and blending colors between two images. I started by defining corresponding points between two faces and used triangulation to create a smooth transition. I then implemented affine transformations to compute a mid-way face, which combines the shapes and colors of the input images. Additionally, I created a morph sequence between two faces, generating an animation from one face to another. Finally, I computed the mean face of a population and extrapolated from this average to produce a caricature. These exercises helped me deepen my understanding of affine warps, image interpolation, and facial feature manipulation.

To define pairs of corresponding points between two faces, I used a labeling tool from a previous CS 180 student. The keypoints were consistently labeled across both faces, ensuring smooth transitions. For this project, I labeled my own face and that of Priyanka Chopra, a well-known Bollywood actress. Once labeled, I computed a Delaunay triangulation based on the midway shape between the two point sets, as this reduces triangle deformation during morphing. This triangulation was used to control the warping process for a smoother morph result.

To create the mid-way face between my face and Priyanka Chopra's, I started by computing the average shape, which involved averaging the corresponding keypoints of both faces. Next, I warped both faces into this new shape by calculating affine transformations for each triangle in the triangulation. These transformations mapped the original triangles to their positions in the average shape. I then used inverse warping to determine the pixel colors for each triangle. Finally, I blended the two images by averaging the colors together, producing a smooth transition. Below, you can see the original images of myself and Priyanka Chopra, along with the resulting mid-way face.

For the morph sequence, I wrote a function that produces a smooth transition between my face and Priyanka Chopra's face. This function uses point correspondences and triangulation to warp both images into intermediate shapes, controlled by parameters for warping and dissolving the images. By varying these parameters, I generated a sequence of frames that smoothly morph my face into Priyanka's. The result is a video sequence as an animated gif.

The results of warping the average geometry to my face, and vice versa, were not ideal. My round and wide face shape does not align well with the average, which led to issues like the chin of the neutral male and female being noticeably cut off.

To create caricatures of my face, I extrapolated from the population mean of the neutral females from the Danes dataset. I calculated the difference between my facial keypoints and the average keypoints, then amplified this difference to create exaggerated versions of my face. By adding 1.5 times the difference, I generated a more pronounced caricature, and with 0.5 times the difference, I created a subtler version. After warping my face into these new shapes, I cropped the images to remove any excess background, highlighting the exaggerated features.