Mercurial > cortex

comparison org/ullman.org @ 379:f1b8727360fb

Find changesets by keywords (author, files, the commit message), revision number or hash, or revset expression.
add images.
author rlm
date Wed, 10 Apr 2013 16:38:52 -0400
parents
children 2d0afb231081
comparison
equal deleted inserted replaced
378:8e62bf52be59 379:f1b8727360fb
1 #+title: Ullman Literature Review
2 #+author: Robert McIntyre
3 #+email: rlm@mit.edu
4 #+description: Review of some of the AI works of Professor Shimon Ullman.
5 #+keywords: Shimon, Ullman, computer vision, artificial intelligence, literature review
6 #+SETUPFILE: ../../aurellem/org/setup.org
7 #+INCLUDE: ../../aurellem/org/level-0.org
8 #+babel: :mkdirp yes :noweb yes :exports both
9
10
11 * Ullman
12
13 Actual code reuse!
14
15 precision = fraction of retrieved instances that are relevant
16 (true-postives/(true-positives+false-positives))
17
18 recall = fraction of relevant instances that are retrieved
19 (true-positives/total-in-class)
20
21 cross-validation = train the model on two different sets to prevent
22 overfitting.
23
24 nifty, relevant, realistic ideas
25 He doesn't confine himself to unplasaubile assumptions
26
27 ** Our Reading
28
29 *** 2002 Visual features of intermediate complexity and their use in classification
30
31
32
33
34 ** Getting around the dumb "fixed training set" methods
35
36 *** 2006 Learning to classify by ongoing feature selection
37
38 Brings in the most informative features of a class, based on
39 mutual information between that feature and all the examples
40 encountered so far. To bound the running time, he uses only a
41 fixed number of the most recent examples. He uses a replacement
42 strategy to tell whether a new feature is better than one of the
43 corrent features.
44
45 *** 2009 Learning model complexity in an online environment
46
47 Sort of like the heirichal baysean models of Tennanbaum, this
48 system makes the model more and more complicated as it gets more
49 and more training data. It does this by using two systems in
50 parallell and then whenever the more complex one seems to be
51 needed by the data, the less complex one is thrown out, and an
52 even more complex model is initialized in its place.
53
54 He uses a SVM with polynominal kernels of varying complexity. He
55 gets good perfoemance on a handwriting classfication using a large
56 range of training samples, since his model changes complexity
57 depending on the number of training samples. The simpler models do
58 better with few training points, and the more complex ones do
59 better with many training points.
60
61 The final model had intermediate complexity between published
62 extremes.
63
64 The more complex models must be able to be initialized efficiently
65 from the less complex models which they replace!
66
67
68 ** Non Parametric Models
69
70 [[../images/viola-parzen-1.png]]
71 [[../images/viola-parzen-2.png]]
72
73 *** 2010 The chains model for detecting parts by their context
74
75 Like the constelation method for rigid objects, but extended to
76 non-rigid objects as well.
77
78 Allows you to build a hand detector from a face detector. This is
79 usefull because hands might be only a few pixels, and very
80 ambiguous in an image, but if you are expecting them at the end of
81 an arm, then they become easier to find.
82
83 They make chains by using spatial proximity of features. That way,
84 a hand can be idntified by chaining back from the head. If there
85 is a good chain to the head, then it is more likely that there is
86 a hand than if there isn't. Since there is some give in the
87 proximity detection, the system can accomodate new poses that it
88 has never seen before.
89
90 Does not use any motion information.
91
92 *** 2005 A Hierarchical Non-Parametric Method for Capturing Non-Rigid Deformations
93
94 (relative dynamic programming [RDP])
95
96 Goal is to match images, as in SIFT, but this time the images can
97 be subject to non rigid transformations. They do this by finding
98 small patches that look the same, then building up bigger
99 patches. They get a tree of patches that describes each image, and
100 find the edit distance between each tree. Editing operations
101 involve a coherent shift of features, so they can accomodate local
102 shifts of patches in any direction. They get some cool results
103 over just straight correlation. Basically, they made an image
104 comparor that is resistant to multiple independent deformations.
105
106 !important small regions are treated the same as nonimportant
107 small regions
108
109 !no conception of shape
110
111 quote:
112 The dynamic programming procedure looks for an optimal
113 transformation that aligns the patches of both images. This
114 transformation is not a global transformation, but a composition
115 of many local transformations of sub-patches at various sizes,
116 performed one on top of the other.
117
118 *** 2006 Satellite Features for the Classification of Visually Similar Classes
119
120 Finds features that can distinguish subclasses of a class, by
121 first finding a rigid set of anghor features that are common to
122 both subclasses, then finding distinguishing features relative to
123 those subfeatures. They keep things rigid because the satellite
124 features don't have much information in and of themselves, and are
125 only informative relative to other features.
126
127 *** 2005 Learning a novel class from a single example by cross-generalization.
128
129 Let's you use a vast visual experience to generate a classifier
130 for a novel class by generating synthetic examples by replaceing
131 features from the single example with features from similiar
132 classes.
133
134 quote: feature F is likely to be useful for class C if a similar
135 feature F proved effective for a similar class C in the past.
136
137 Allows you to trasfer the "gestalt" of a similiar class to a new
138 class, by adapting all the features of the learned class that have
139 correspondance to the new class.
140
141 *** 2007 Semantic Hierarchies for Recognizing Objects and Parts
142
143 Better learning of complex objects like faces by learning each
144 piece (like nose, mouth, eye, etc) separately, then making sure
145 that the features are in plausable positions.