Multiscale tricks If it must fail, it must fail early Iterative Reduction by (re-)Weighting
Statistical Model of Shape & Appearance for constrained search [Tim Cootes] Shape : point location Appearance : image patch around point X = Xavg + StdDev X = Xavg + Pb P : PCA , b : parameter (reduced search space) Steps : Foreach point in shape model: ▪ Find most similar patches along normal direction ▪ Test if current shape fit with learned model ▪ Continue until convergence
Hx = z, given x & z, find H Set of pair of matched points Overdetermined system of linear equation RANSAC Pick small subset Calculate H Check whether H is supported by the rest of population Select best H with biggest support (minimize error)
Fill patches with highest information content Structure (edge) Boundary-patch High variance : chaotic textures (e.g. grass) Low variance : plain
Math can also means Fun especially when it’s applied, you don’t know what you’ve missed Innovation in Research seeks the balance between pragmatic and platonic (practical vs ideal), there are still lot of areas to explore For everything one can imagine, someone is making it somewhere
DO NOT Afraid! DO NOT Worry! DO NOT Hesitate!