annotate org/ullman.org @ 406:40b67bb71430

save progress.
author Robert McIntyre <rlm@mit.edu>
date Tue, 18 Mar 2014 18:34:10 -0400
parents 2d0afb231081
children
rev   line source
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@380 16 (true-positives/(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@380 22 overfitting, and confirm that you have enough training samples.
rlm@379 23
rlm@379 24 nifty, relevant, realistic ideas
rlm@380 25 He doesn't confine himself to implausible 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@380 34 Viola's PhD thesis has a good introduction to entropy and mutual
rlm@380 35 information
rlm@380 36
rlm@379 37 ** Getting around the dumb "fixed training set" methods
rlm@379 38
rlm@379 39 *** 2006 Learning to classify by ongoing feature selection
rlm@379 40
rlm@379 41 Brings in the most informative features of a class, based on
rlm@379 42 mutual information between that feature and all the examples
rlm@379 43 encountered so far. To bound the running time, he uses only a
rlm@379 44 fixed number of the most recent examples. He uses a replacement
rlm@379 45 strategy to tell whether a new feature is better than one of the
rlm@380 46 current features.
rlm@379 47
rlm@379 48 *** 2009 Learning model complexity in an online environment
rlm@379 49
rlm@380 50 Sort of like the hierarchical Bayesan models of Tennanbaum, this
rlm@379 51 system makes the model more and more complicated as it gets more
rlm@379 52 and more training data. It does this by using two systems in
rlm@380 53 parallel and then whenever the more complex one seems to be
rlm@379 54 needed by the data, the less complex one is thrown out, and an
rlm@379 55 even more complex model is initialized in its place.
rlm@379 56
rlm@380 57 He uses a SVM with polynomial kernels of varying complexity. He
rlm@380 58 gets good performance on a handwriting classification using a large
rlm@379 59 range of training samples, since his model changes complexity
rlm@379 60 depending on the number of training samples. The simpler models do
rlm@379 61 better with few training points, and the more complex ones do
rlm@379 62 better with many training points.
rlm@379 63
rlm@379 64 The final model had intermediate complexity between published
rlm@379 65 extremes.
rlm@379 66
rlm@379 67 The more complex models must be able to be initialized efficiently
rlm@379 68 from the less complex models which they replace!
rlm@379 69
rlm@379 70
rlm@379 71 ** Non Parametric Models
rlm@379 72
rlm@379 73 [[../images/viola-parzen-1.png]]
rlm@379 74 [[../images/viola-parzen-2.png]]
rlm@379 75
rlm@379 76 *** 2010 The chains model for detecting parts by their context
rlm@379 77
rlm@380 78 Like the constellation method for rigid objects, but extended to
rlm@379 79 non-rigid objects as well.
rlm@379 80
rlm@379 81 Allows you to build a hand detector from a face detector. This is
rlm@380 82 useful because hands might be only a few pixels, and very
rlm@379 83 ambiguous in an image, but if you are expecting them at the end of
rlm@379 84 an arm, then they become easier to find.
rlm@379 85
rlm@379 86 They make chains by using spatial proximity of features. That way,
rlm@380 87 a hand can be identified by chaining back from the head. If there
rlm@379 88 is a good chain to the head, then it is more likely that there is
rlm@379 89 a hand than if there isn't. Since there is some give in the
rlm@380 90 proximity detection, the system can accommodate new poses that it
rlm@379 91 has never seen before.
rlm@379 92
rlm@379 93 Does not use any motion information.
rlm@379 94
rlm@379 95 *** 2005 A Hierarchical Non-Parametric Method for Capturing Non-Rigid Deformations
rlm@379 96
rlm@379 97 (relative dynamic programming [RDP])
rlm@379 98
rlm@379 99 Goal is to match images, as in SIFT, but this time the images can
rlm@379 100 be subject to non rigid transformations. They do this by finding
rlm@379 101 small patches that look the same, then building up bigger
rlm@379 102 patches. They get a tree of patches that describes each image, and
rlm@379 103 find the edit distance between each tree. Editing operations
rlm@380 104 involve a coherent shift of features, so they can accommodate local
rlm@379 105 shifts of patches in any direction. They get some cool results
rlm@379 106 over just straight correlation. Basically, they made an image
rlm@380 107 comparator that is resistant to multiple independent deformations.
rlm@379 108
rlm@380 109 !important small regions are treated the same as unimportant
rlm@379 110 small regions
rlm@379 111
rlm@379 112 !no conception of shape
rlm@379 113
rlm@379 114 quote:
rlm@379 115 The dynamic programming procedure looks for an optimal
rlm@379 116 transformation that aligns the patches of both images. This
rlm@379 117 transformation is not a global transformation, but a composition
rlm@379 118 of many local transformations of sub-patches at various sizes,
rlm@379 119 performed one on top of the other.
rlm@379 120
rlm@379 121 *** 2006 Satellite Features for the Classification of Visually Similar Classes
rlm@379 122
rlm@379 123 Finds features that can distinguish subclasses of a class, by
rlm@380 124 first finding a rigid set of anchor features that are common to
rlm@379 125 both subclasses, then finding distinguishing features relative to
rlm@379 126 those subfeatures. They keep things rigid because the satellite
rlm@379 127 features don't have much information in and of themselves, and are
rlm@379 128 only informative relative to other features.
rlm@379 129
rlm@379 130 *** 2005 Learning a novel class from a single example by cross-generalization.
rlm@379 131
rlm@379 132 Let's you use a vast visual experience to generate a classifier
rlm@380 133 for a novel class by generating synthetic examples by replacing
rlm@380 134 features from the single example with features from similar
rlm@379 135 classes.
rlm@379 136
rlm@379 137 quote: feature F is likely to be useful for class C if a similar
rlm@379 138 feature F proved effective for a similar class C in the past.
rlm@379 139
rlm@380 140 Allows you to transfer the "gestalt" of a similar class to a new
rlm@379 141 class, by adapting all the features of the learned class that have
rlm@380 142 correspondence to the new class.
rlm@379 143
rlm@379 144 *** 2007 Semantic Hierarchies for Recognizing Objects and Parts
rlm@379 145
rlm@379 146 Better learning of complex objects like faces by learning each
rlm@379 147 piece (like nose, mouth, eye, etc) separately, then making sure
rlm@380 148 that the features are in plausible positions.