[Computer Vision 4] Distinctive Image Features from Scale-Invariant Keypoints

Detection of Scale-space Extrema

Detection location that are invariant to scale: search for stable features across all possible scales.

L(x,y,σ) = G(x,y,σ) ∗ I(x,y): L function of scale space of image, G a variable scale Gaussian, I input image

Difference of Gaussian: s: interval in DOG, s+2 layers need for extrema finding process, s+3 layers need for Gaussian, k = 2^(1/s)

Find local extrema -> Determine the accuracy location by interpolation -> Eliminate low contrast points -> Eliminate points on edges (using Hessian Matrix)

Local Image Descriptor

rotation invariant, invariant to affine changes in illumination

Descriptor vector size: r^n^2  r: number of orientation histogram bins, n: size of grids