100 likes | 257 Views
Report 1: Optical Flow and Sift. Billy Timlen. Lucas Kanade. ( u,v ) = inv(A t A)* A t *F t Derived from f x *u +f y *v = -f t (after taking the partial derivative in terms of each variable x,y,t Analyze the pixels around the point of interest Requires a degree of padding
E N D
Report 1: Optical Flow and Sift Billy Timlen
Lucas Kanade • (u,v) = inv(AtA)*At*Ft • Derived from fx*u +fy*v = -ft (after taking the partial derivative in terms of each variable x,y,t • Analyze the pixels around the point of interest • Requires a degree of padding • Works for slow motion and small areas
Optical Flow with Gaussian Pyramids • Reduces the original image into different levels • Impyramid(image, ‘reduce’) • Computes Optical flow for each level • Shifts derivative mask by u and v of prior level • Add the optical flows of each level • Should record more detailed results of motion
Sift • Input: 18x18 patch, keypoint and orientation angle • Outputs a descriptor • Histogram of orientation magnitudes • Results vary according to the Gaussian used (for smoothing) and the sigma used (which affects the Gaussian)
What Next? • Work with different types of masks • Use different forms of interpolation • MatLab has their own function • Use another form of rounding the non-integer indices from u and v • Gonzalo sent us a bilinear function to look at
Possible Projects • Optimal Algorithms for Topologically Constrained Correspondence • Bayesian Formulation for Event Recounting given Event Label • 3D Joint Localization for Gesture Recognition • GPS-Tag Refinement