Race Training Data

---

Training faces from each class were projected onto the 200 eigenvectors of the complete training set with largest eigenvalues.

A Fisher Projection was then used to reduce the 200-dimensional data to three dimensions.

 

Fisher Projected Training Data for Asian, Black, and Hispanic

[larger version]

 

Fisher Projected Training Data for Asian, Black, and Hispanic, and White

[larger version]

 

Alternative View of Complete Training Data

[larger version]

 

Clearly, the data is not linearly separable. We will need to use a classification technique capable of yielding non-linear decision boundaries.