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  X The  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