15-862 Computational Photography

Project 3: Face Morphing
 by: Lisa Chan

Index:

  1. Face Morphing
    1. Basic Face Morphing
    2. Average Face and Caricatures
    3. Principal Component Analysis and Caricatures

  2. Other Morphing
    1. Car Morphing
    2. Bottle Morphing
    3. Transforming Toy

Project Details:

To have the images look better during the warp, additional points were added to the image. On each side of the image, 10 correspondence points were added. On the first trial of computing the warp, the program took 1 hour to produce 1 frame. The reason was because interp2 was actually really slow at interpolating each point individually. Once interp2 was taken outside the for loops and was fed vectors instead of points, each frame only took 26 seconds.


1. Face Morphing:

A. Basic Face Morphing:

PictureA
PictureB
Morphing Gif


B. Average Face & Caricatures:

The average face of the class clearly shows that there are more males than females!! The average male face looks almost identical to the average face of the class. Since the TA and the professor made funny faces, an average face without their images was calculated. There is no difference between the two average faces, mainly because there were more “normal” people than the two outliers of the data set. Also, the faint outline of glasses indicate that the majority of the class wears glasses.

Average Face of Students
Average Face of Whole Data Set
Average Male Face
Average Female Face


Before performing actual calculations of caricatures, simply looking at everyone’s face morphed to the average geometry is in fact looking at caricatures!

Original Image
Morphed Into Average Geometry

Original Image
Morphed Into Average Geometry

Original Image
Morphed Into Average Geometry


The average faces morphed into my geometry is horrible! Not that my face morphed into the average geometry is any better.

Average Class Face Into My Geometry
Average Male Face Into My Geometry
Average Female Face Into My Geometry
My Face Into Average Geometry


To make my face more female or male, I simply cross dissolved the average face in my geometry with my original image. However, since the average face in my geometry doesn't look very good, the results also too terrible. The numbers indicated are the cross dissolve fraction used for my face.

0.2
0.3
0.4
0.5
0.6

0.2
0.3
0.4
0.5
0.6


Here are the results of me morphed into the average female or male face. In general, the pictures don't look very good because we only selected corresponding points for the face. If we had included the shoulders and the long hair, the results might look better. Also, there are simply not enough data for the girls! Click on the images to see animaed gifs!

Me Morphed into Average Female
Me Morphed into Average Male


Here are the caricatures. I simply warped my face into a geometry that is several times the greater than the corresponding points for the average face. The images are displayed in the order of increasing weighting of the average face when calculating the triangulation.

Using Points From Average Female Face
1.5*avg
2*avg
2.5*avg
3*avg
3.5*avg

Using Points From Average Male Face
1.5*avg
2*avg
2.5*avg
3*avg
3.5*avg

C. Principal Component Analysis & Caricatures:

The PCA images were reconstructed with the eigenvectors to see which component is the most important in measuring the average faces. The first eigenvector is actually the average face. The second to fourth ones seem to only focus on the lighting of the shirt. Interestingly, it seems that the 10th eigenvector highlights the glasses.



I don't exactly understand how to perform a PCA on the shape distortions between the base images and their average warps, so I was only able to vary the colors of the images by changing values in S, V, and U for what I've calculated.



2. Other Morphing:

A. Car Morphing:

Cars don’t really lie in the same subspace, but since the sides of the cars almost always looks the same…with two wheels, windows, headlights, and taillights, the morphing turned out nicely. Of course, I only selected sedans so that there won’t be too much of a variation between the bodies of the cars.

65 Corresponding Points
Triangulation
Morphing Gif

Honda FCX Concept
Accord Coupe
Accord Hybrid Sedan
Accord Sedan
Civic Hybrid Sedan
Civic GX NGV
Kia Optima
Kia Spectra
Kia Rio

B. Bottle Morphing:

The bottle morphing idea came from Professor Efros, he stated that bottles are in the same subspace, but not the lettering on the bottles. However, by selecting the correspondence points so that it marks where the label starts and ends really helps the morphing to look smooth.

30 Corresponding Points
Triangulation
Morphing Gif


C. Transforming Plane:

This was Begum's idea, of morphing images taken at different transforming steps for a toy. Unfortunately, parts of the toy that used to be behind the toy were rotated out in the next step. The morphing sequence then got all screwed up...

102 Corresponding Points
Triangulation
Morphing Gif



3. Image Sources:

  1. Car Morphing
  2. Bottle Morphing