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