Overview: Goal: Face alignment in unseen images. Closely related to Constrained Local Models (CLM) and Ac)ve Shape Models (ASM), where a set of local detectors is constrained to lie in the subspace spanned by a Point Distribu)on Model (PDM). Two step fiOng approach: Two step fiOng approach: (1) Local search using the local detectors (response maps for each landmark). (2) Global op)miza)on strategy that finds the PDM parameters that jointly maximize all the detec)on at once. • New Bayesian global op)miza)on strategy using second order sta)s)cs of the shape and pose parameters. The Shape (PDM) and Appearance Models Local Detectors (MOSSE Filters) Correla)on in Fourier Domain PMOSSE Filter XThe Alignment Goal Given a shape observa)on (y), find the op)mal set of shape (b) and pose parameters that maximize the posterior probability Assuming: Assuming: ① Condi)onal independence between landmarks ② Close to a solu)on The Likelihood Term The Prior Term Local Op)miza)on Strategies (Finding the Likelihood Parameters)Weighted Peak Response (WPR) (current mesh es)mate) Gaussian Response (GR) Kernel Density Es)mator (KDE) 2nd Order MAP Global Alignment (DBASM) Qualita)ve Results -‐ Labeled Faces in the Wild Evalua)ng Global Op)miza)on Strategies Tracking Performance -‐ FGNET Talking Face Sequence