1. CS 332 Visual Processing in Computer and Biological Vision Systems
Hodgepodge

2. Artificial Neural Nets
Use feedforward network to compute output from inputs
Use back-propagation algorithm to learn weights from training data (correct input/output pairs)

3. Rowley, Baluja & Kanade: face detection with neural nets

4. Edwin Land’s color mondrian experiments
Analysis of color
Edwin Land’s color mondrian experiments

5. Land’s Retinex Theory of Color
L(x,y,) = I(x,y,) * R(x,y,)
L(x,y,): luminance
I(x,y,): illuminant
R(x,y,): surface reflectance
Goal: recover surface reflectance (color)

6. Measuring color by retinal cones

7. Principal components analysis
 Method for reducing the dimensionality of a high-dimensional data set, allowing a more compact representation of each element of the set
 Takes advantage of redundancy within a data set
Expresses original data samples as a linear combination of a set of components that capture as much as possible of the data’s variance
Mathematically, the principle components are the eigenvectors of the covariance matrix of the original data set

8. Troje: Using PCA to represent human gait
Obtain motion capture data from many human walkers
 Use PCA to construct a small number of “eigenpostures”
 Express each posture in the original motion sequence as a weighted sum of eigenpostures
Use pattern of changing coefficients over time to recognize movements, e.g. classify gender
Troje walker demo

9. Using eigenpostures to represent gaits
Each posture consists of (x,y,z) coordinates of 15 locations
Each sequence consists of about 1400 postures (12 secs, 20 steps)
PCA analysis: first component captures 84% of variance in postures, first four components capture 98% of variance
P = P0 + ΣciPi
P0: average posture
Pi: i-th principal component, i = 1..4 (four eigenpostures)
Ci: coefficient of i-th eigenposture
Pattern of coefficients over time is approximately sinusoidal – demo varies properties of sinusoids