• Summary
• Contributors
• HOG (Histogram of Oriented Gradients)
  • Original Scale
  • 1/2x Scale
  • 2x Scale
• SVM (Support Vector Machine)
• NMS (Non Maximum Suppression)
• Results

Summary

Multi-scale Face Detection using SVM and Histogram of Oriented Gradients features.

Contributors

• Adam Sorrenti
• Eisa Keramati nejad

HOG (Histogram of Oriented Gradients)

Consider matrix M, a 3x3 image segment representing pixel intensity(I), where (r,c) is a given pixel.

        [ 254 143 22 ] 
    M = [ 230 150 25 ]  
        [ 253 154 21 ] 

Calculate the gradient of the image: $$G_x(r,c)=I(r,c+1)-I(r,c-1)$$ $$G_y(r,c)=I(r-1,c)-I(r+1,c)$$

From left to right shows, an intensity image, $G_x$, and $G_y$

Calculate gradient direction and magnitude: $$Magnitude(µ)=\sqrt{G_x^2+G_y^2}$$ $$Angle(θ)=|tan^{-1}(G_y/G_x)|$$

Compute HOG:

Original Scale

1/2x Scale

2x Scale

SVM (Support Vector Machine)

Faces training set: (36x36 pixel images)

Not faces set:

NMS (Non Maximum Suppression)

Results