Mercurial > cortex
comparison org/ullman.org @ 382:9b3487a515a7
merge
author | Robert McIntyre <rlm@mit.edu> |
---|---|
date | Tue, 16 Apr 2013 03:51:41 +0000 |
parents | 2d0afb231081 |
children |
comparison
equal
deleted
inserted
replaced
381:9ac42f1fdf0a | 382:9b3487a515a7 |
---|---|
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-positives/(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, and confirm that you have enough training samples. | |
23 | |
24 nifty, relevant, realistic ideas | |
25 He doesn't confine himself to implausible assumptions | |
26 | |
27 ** Our Reading | |
28 | |
29 *** 2002 Visual features of intermediate complexity and their use in classification | |
30 | |
31 | |
32 | |
33 | |
34 Viola's PhD thesis has a good introduction to entropy and mutual | |
35 information | |
36 | |
37 ** Getting around the dumb "fixed training set" methods | |
38 | |
39 *** 2006 Learning to classify by ongoing feature selection | |
40 | |
41 Brings in the most informative features of a class, based on | |
42 mutual information between that feature and all the examples | |
43 encountered so far. To bound the running time, he uses only a | |
44 fixed number of the most recent examples. He uses a replacement | |
45 strategy to tell whether a new feature is better than one of the | |
46 current features. | |
47 | |
48 *** 2009 Learning model complexity in an online environment | |
49 | |
50 Sort of like the hierarchical Bayesan models of Tennanbaum, this | |
51 system makes the model more and more complicated as it gets more | |
52 and more training data. It does this by using two systems in | |
53 parallel and then whenever the more complex one seems to be | |
54 needed by the data, the less complex one is thrown out, and an | |
55 even more complex model is initialized in its place. | |
56 | |
57 He uses a SVM with polynomial kernels of varying complexity. He | |
58 gets good performance on a handwriting classification using a large | |
59 range of training samples, since his model changes complexity | |
60 depending on the number of training samples. The simpler models do | |
61 better with few training points, and the more complex ones do | |
62 better with many training points. | |
63 | |
64 The final model had intermediate complexity between published | |
65 extremes. | |
66 | |
67 The more complex models must be able to be initialized efficiently | |
68 from the less complex models which they replace! | |
69 | |
70 | |
71 ** Non Parametric Models | |
72 | |
73 [[../images/viola-parzen-1.png]] | |
74 [[../images/viola-parzen-2.png]] | |
75 | |
76 *** 2010 The chains model for detecting parts by their context | |
77 | |
78 Like the constellation method for rigid objects, but extended to | |
79 non-rigid objects as well. | |
80 | |
81 Allows you to build a hand detector from a face detector. This is | |
82 useful because hands might be only a few pixels, and very | |
83 ambiguous in an image, but if you are expecting them at the end of | |
84 an arm, then they become easier to find. | |
85 | |
86 They make chains by using spatial proximity of features. That way, | |
87 a hand can be identified by chaining back from the head. If there | |
88 is a good chain to the head, then it is more likely that there is | |
89 a hand than if there isn't. Since there is some give in the | |
90 proximity detection, the system can accommodate new poses that it | |
91 has never seen before. | |
92 | |
93 Does not use any motion information. | |
94 | |
95 *** 2005 A Hierarchical Non-Parametric Method for Capturing Non-Rigid Deformations | |
96 | |
97 (relative dynamic programming [RDP]) | |
98 | |
99 Goal is to match images, as in SIFT, but this time the images can | |
100 be subject to non rigid transformations. They do this by finding | |
101 small patches that look the same, then building up bigger | |
102 patches. They get a tree of patches that describes each image, and | |
103 find the edit distance between each tree. Editing operations | |
104 involve a coherent shift of features, so they can accommodate local | |
105 shifts of patches in any direction. They get some cool results | |
106 over just straight correlation. Basically, they made an image | |
107 comparator that is resistant to multiple independent deformations. | |
108 | |
109 !important small regions are treated the same as unimportant | |
110 small regions | |
111 | |
112 !no conception of shape | |
113 | |
114 quote: | |
115 The dynamic programming procedure looks for an optimal | |
116 transformation that aligns the patches of both images. This | |
117 transformation is not a global transformation, but a composition | |
118 of many local transformations of sub-patches at various sizes, | |
119 performed one on top of the other. | |
120 | |
121 *** 2006 Satellite Features for the Classification of Visually Similar Classes | |
122 | |
123 Finds features that can distinguish subclasses of a class, by | |
124 first finding a rigid set of anchor features that are common to | |
125 both subclasses, then finding distinguishing features relative to | |
126 those subfeatures. They keep things rigid because the satellite | |
127 features don't have much information in and of themselves, and are | |
128 only informative relative to other features. | |
129 | |
130 *** 2005 Learning a novel class from a single example by cross-generalization. | |
131 | |
132 Let's you use a vast visual experience to generate a classifier | |
133 for a novel class by generating synthetic examples by replacing | |
134 features from the single example with features from similar | |
135 classes. | |
136 | |
137 quote: feature F is likely to be useful for class C if a similar | |
138 feature F proved effective for a similar class C in the past. | |
139 | |
140 Allows you to transfer the "gestalt" of a similar class to a new | |
141 class, by adapting all the features of the learned class that have | |
142 correspondence to the new class. | |
143 | |
144 *** 2007 Semantic Hierarchies for Recognizing Objects and Parts | |
145 | |
146 Better learning of complex objects like faces by learning each | |
147 piece (like nose, mouth, eye, etc) separately, then making sure | |
148 that the features are in plausible positions. |