diff org/literature-review.org @ 377:80cd096682b2

reviewing ullman's stuff.
author Robert McIntyre <rlm@mit.edu>
date Thu, 11 Apr 2013 06:19:59 +0000
parents 057d47fc4789
children 8e62bf52be59
line wrap: on
line diff
     1.1 --- a/org/literature-review.org	Thu Apr 11 05:40:23 2013 +0000
     1.2 +++ b/org/literature-review.org	Thu Apr 11 06:19:59 2013 +0000
     1.3 @@ -208,6 +208,8 @@
     1.4  cross-validation = train the model on two different sets to prevent
     1.5  overfitting. 
     1.6  
     1.7 +nifty, relevant, realistic ideas
     1.8 +He doesn't confine himself to unplasaubile assumptions
     1.9  
    1.10  
    1.11  
    1.12 @@ -239,15 +241,20 @@
    1.13      better with few training points, and the more complex ones do
    1.14      better with many training points.
    1.15  
    1.16 +    The final model had intermediate complexity between published
    1.17 +    extremes. 
    1.18 +
    1.19      The more complex models must be able to be initialized efficiently
    1.20      from the less complex models which they replace!
    1.21  
    1.22  
    1.23  ** Non Parametric Models
    1.24  
    1.25 -*** Visual features of intermediate complexity and their use in classification
    1.26 +*** 2002 Visual features of intermediate complexity and their use in classification
    1.27  
    1.28 -*** The chains model for detecting parts by their context
    1.29 +    
    1.30 +
    1.31 +*** 2010 The chains model for detecting parts by their context
    1.32  
    1.33      Like the constelation method for rigid objects, but extended to
    1.34      non-rigid objects as well.
    1.35 @@ -257,4 +264,60 @@
    1.36      ambiguous in an image, but if you are expecting them at the end of
    1.37      an arm, then they become easier to find.
    1.38  
    1.39 -    
    1.40 \ No newline at end of file
    1.41 +    They make chains by using spatial proximity of features. That way,
    1.42 +    a hand can be idntified by chaining back from the head. If there
    1.43 +    is a good chain to the head, then it is more likely that there is
    1.44 +    a hand than if there isn't. Since there is some give in the
    1.45 +    proximity detection, the system can accomodate new poses that it
    1.46 +    has never seen before.
    1.47 +
    1.48 +    Does not use any motion information.
    1.49 +
    1.50 +*** 2005 A Hierarchical Non-Parametric Method for Capturing Non-Rigid Deformations
    1.51 +    
    1.52 +    (relative dynamic programming [RDP])
    1.53 +
    1.54 +    Goal is to match images, as in SIFT, but this time the images can
    1.55 +    be subject to non rigid transformations. They do this by finding
    1.56 +    small patches that look the same, then building up bigger
    1.57 +    patches. They get a tree of patches that describes each image, and
    1.58 +    find the edit distance between each tree. Editing operations
    1.59 +    involve a coherent shift of features, so they can accomodate local
    1.60 +    shifts of patches in any direction. They get some cool results
    1.61 +    over just straight correlation. Basically, they made an image
    1.62 +    comparor that is resistant to multiple independent deformations.
    1.63 +    
    1.64 +    !important small regions are treated the same as nonimportant
    1.65 +     small regions
    1.66 +     
    1.67 +    !no conception of shape
    1.68 +    
    1.69 +    quote:
    1.70 +    The dynamic programming procedure looks for an optimal
    1.71 +    transformation that aligns the patches of both images. This
    1.72 +    transformation is not a global transformation, but a composition
    1.73 +    of many local transformations of sub-patches at various sizes,
    1.74 +    performed one on top of the other.
    1.75 +
    1.76 +*** 2006 Satellite Features for the Classification of Visually Similar Classes
    1.77 +    
    1.78 +    Finds features that can distinguish subclasses of a class, by
    1.79 +    first finding a rigid set of anghor features that are common to
    1.80 +    both subclasses, then finding distinguishing features relative to
    1.81 +    those subfeatures. They keep things rigid because the satellite
    1.82 +    features don't have much information in and of themselves, and are
    1.83 +    only informative relative to other features.
    1.84 +
    1.85 +*** 2005 Learning a novel class from a single example by cross-generalization.
    1.86 +
    1.87 +    Let's you use a vast visual experience to generate a classifier
    1.88 +    for a novel class by generating synthetic examples by replaceing
    1.89 +    features from the single example with features from similiar
    1.90 +    classes.
    1.91 +
    1.92 +    quote: feature F is likely to be useful for class C if a similar
    1.93 +    feature F proved effective for a similar class C in the past.
    1.94 +
    1.95 +    Allows you to trasfer the "gestalt" of a similiar class to a new
    1.96 +    class, by adapting all the features of the learned class that have
    1.97 +    correspondance to the new class.
    1.98 \ No newline at end of file