Learning in Computer Vision Simon Lucey
<p>Dec 08, 2010 03:29 PM 6 part Lecture:  quite understandable
 
Simon Lucey  CMU video  (boss is Takeo Kanade  tah KAY oh  Cah NAH day)
 
big problem is REGISTRATION  (registering the model with the image?)
 
from  Machine Learning Summer School 2008 in Kioloa
 
face recognition:  where is it?  recog an expression
til 5 years ago: just putting a box around the face:
 
pose estimation:  frontal face:  interest point detectors:  then  map those onto a cylinder
 
body tracking  nonrigid mvmt;  
 
visual speech recog  (he did his PhD thesis on that yrs ago)
 
palm recognition  same techniques as head
 
car tracking
 
skip th outline and go to 14:30  Task of Alighnment and REgistration
 
I want to know where the eyes/nose are in the picture.
 
warping function is a mapping function   (affine means  translation, rotation, scaling)
(non-rigid warps are hard: like towels)
 
illumination causes problems;  so does changed facial expression
want to register an object even when expression and pose varies   minute 17
 
Kaneda says REGISTRATION is the key;
 
Learning: how do I lean an object template/model that satsifactorily discriminate btwn the
object and the image background
 
Fitting:  how do I effriciently evaluate the object template/model across all possible warp values.
 
minute 24: lots of variations in an image:  lighting;  pose; occlusions (hands over face);  noise/nuisance variables;
occlusion:  person walks behind a house;  hair over the face; hats etc.
 
min 26:  even just 640 by 480 image:  hellish
to find correct scale... requires  24,000 windows.
 
WARP functions  W(x; p)   x is the coordinate point and p is the parametric form
 
z us tge concatenation (vector) of all the points in the template.
 
M = pseudoinverse of ....  *  X
 
M*src = dst    src = source  dst = destination
 
translation and affine are the basic warps.  (affine also includes shear)
 
aliasing  (wagon wheels going in wrong direction; bad sampling)
can do interpolation  eg bicubic; eg nyquist... sample at twice the high frequency.
can do bicubic or bilinear
 
take away mean intensity;  do histogram equalization (to adjust brightness)
 
How to sample the warp parameters?
might do every tranlation; but cannot do every affine transform etc.
 
SSD = sum of squared differences;  altho not always perfet,
common measure of image similarity is  SSD  (but susceptible to outliers)
 
corr(p)  dot product
 
 
 
 
 
 
 </p>