1.1 --- a/org/literature-review.org	Tue Mar 12 04:34:37 2013 +0000
     1.2 +++ b/org/literature-review.org	Thu Apr 11 05:40:23 2013 +0000
     1.3 @@ -42,13 +42,13 @@
     1.4  
     1.5   - Zhang, Z., R. Deriche, O. Faugeras, Q.T. Luong, “A robust
     1.6    technique for matching two uncalibrated images through the recovery
     1.7 -  of the unknown epipolar geometry,” Artificial In- telligence, 78,
     1.8 +  of the unknown epipolar geometry,” Artificial Intelligence, 78,
     1.9    (1995), pp. 87-119.
    1.10  
    1.11  
    1.12  
    1.13  
    1.14 -
    1.15 +   
    1.16  * Alignment by Maximization of Mutual Information, Paul A. Viola
    1.17  
    1.18    PhD Thesis recommended by Winston. Describes a system that is able
    1.19 @@ -63,18 +63,198 @@
    1.20    - Differential entropy seems a bit odd -- you would think that it
    1.21      should be the same as normal entropy for a discrete distrubition
    1.22      embedded in continuous space. How do you measure the entropy of a
    1.23 -    half continuous, half discrete random variable?
    1.24 +    half continuous, half discrete random variable? Perhaps the
    1.25 +    problem is related to the delta function, and not the definition
    1.26 +    of differential entropy?
    1.27  
    1.28    - Expectation Maximation (Mixture of Gaussians cool stuff)
    1.29      (Dempster 1977)
    1.30  
    1.31    - Good introduction to Parzen Window Density Estimation. Parzen
    1.32      density functions trade construction time for evaulation
    1.33 -    time.(Pg. 41)
    1.34 +    time.(Pg. 41) They are a way to transform a sample into a
    1.35 +    distribution. They don't work very well in higher dimensions due
    1.36 +    to the thinning of sample points.
    1.37 +
    1.38 +  - Calculating the entropy of a Markov Model (or state machine,
    1.39 +    program, etc) seems like it would be very hard, since each trial
    1.40 +    would not be independent of the other trials. Yet, there are many
    1.41 +    common sense models that do need to have state to accurately model
    1.42 +    the world.
    1.43 +
    1.44 +  - "... there is no direct procedure for evaluating entropy from a
    1.45 +    sample. A common approach is to model the density from the sample,
    1.46 +    and then estimate the entropy from the density."
    1.47 +
    1.48 +  - pg. 55 he says that infinity minus infinity is zero lol.
    1.49 +
    1.50 +  - great idea on pg 62 about using random samples from images to
    1.51 +    speed up computation.
    1.52 +
    1.53 +  - practical way of terminating a random search: "A better idea is to
    1.54 +    reduce the learning rate until the parameters have a reasonable
    1.55 +    variance and then take the average parameters."
    1.56 +
    1.57 +  - p. 65 bullshit hack to make his parzen window estimates work.
    1.58 +
    1.59 +  - this alignment only works if the initial pose is not very far
    1.60 +    off. 
    1.61 +
    1.62  
    1.63    Occlusion? Seems a bit holistic.
    1.64  
    1.65 +** References
    1.66 + - "excellent" book on entropy (Cover & Thomas, 1991) [Elements of
    1.67 +   Information Theory.] 
    1.68 +
    1.69 + - Canny, J. (1986). A Computational Approach to Edge Detection. IEEE
    1.70 +   Transactions PAMI, PAMI-8(6):679{698
    1.71 +
    1.72 + - Chin, R. and Dyer, C. (1986). Model-Based Recognition in Robot
    1.73 +   Vision. Computing Surveys, 18:67-108.
    1.74 +
    1.75 + - Grimson, W., Lozano-Perez, T., Wells, W., et al. (1994). An
    1.76 +   Automatic Registration Method for Frameless Stereotaxy, Image
    1.77 +   Guided Surgery, and Enhanced Realigy Visualization. In Proceedings
    1.78 +   of the Computer Society Conference on Computer Vision and Pattern
    1.79 +   Recognition, Seattle, WA. IEEE.
    1.80 +
    1.81 + - Hill, D. L., Studholme, C., and Hawkes, D. J. (1994). Voxel
    1.82 +   Similarity Measures for Auto-mated Image Registration. In
    1.83 +   Proceedings of the Third Conference on Visualization in Biomedical
    1.84 +   Computing, pages 205 { 216. SPIE.
    1.85 +
    1.86 + - Kirkpatrick, S., Gelatt, C., and Vecch Optimization by Simulated
    1.87 +   Annealing. Science, 220(4598):671-680.
    1.88 +
    1.89 + - Jones, M. and Poggio, T. (1995). Model-based matching of line
    1.90 +   drawings by linear combin-ations of prototypes. Proceedings of the
    1.91 +   International Conference on Computer Vision
    1.92 +
    1.93 + - Ljung, L. and Soderstrom, T. (1983). Theory and Practice of
    1.94 +   Recursive Identi cation. MIT Press.
    1.95 +
    1.96 + - Shannon, C. E. (1948). A mathematical theory of communication. Bell
    1.97 +   Systems Technical Journal, 27:379-423 and 623-656.
    1.98 +
    1.99 + - Shashua, A. (1992). Geometry and Photometry in 3D Visual
   1.100 +   Recognition. PhD thesis, M.I.T Artificial Intelligence Laboratory,
   1.101 +   AI-TR-1401.
   1.102 +
   1.103 + - William H. Press, Brian P. Flannery, S. A. T. and Veterling,
   1.104 +   W. T. (1992). Numerical Recipes in C: The Art of Scienti c
   1.105 +   Computing. Cambridge University Press, Cambridge, England, second
   1.106 +   edition edition.
   1.107 +
   1.108 +* Semi-Automated Dialogue Act Classification for Situated Social Agents in Games, Deb Roy 
   1.109 +  
   1.110 +  Interesting attempt to learn "social scripts" related to resturant
   1.111 +  behaviour. The authors do this by creating a game which implements a
   1.112 +  virtual restruant, and recoding actual human players as they
   1.113 +  interact with the game. The learn scripts from annotated
   1.114 +  interactions and then use those scripts to label other
   1.115 +  interactions. They don't get very good results, but their
   1.116 +  methodology of creating a virtual world and recording
   1.117 +  low-dimensional actions is interesting.
   1.118 +
   1.119 +  - Torque 2D/3D looks like an interesting game engine.
   1.120 +
   1.121 +
   1.122 +* Face Recognition by Humans: Nineteen Results all Computer Vision Researchers should know, Sinha
   1.123 +  
   1.124 +  This is a summary of a lot of bio experiments on human face
   1.125 +  recognition.
   1.126 +  
   1.127 +  - They assert again that the internal gradients/structures of a face
   1.128 +    are more important than the edges.
   1.129 +
   1.130 +  - It's amazing to me that it takes about 10 years after birth for a
   1.131 +    human to get advanced adult-like face detection. They go through
   1.132 +    feature based processing to a holistic based approach during this
   1.133 +    time.
   1.134 +
   1.135 +  - Finally, color is a very important cue for identifying faces.
   1.136  
   1.137  ** References
   1.138 - - "excellent" book on entropy (Cover & Thomas, 1991)
   1.139 - 
   1.140 \ No newline at end of file
   1.141 +  - A. Freire, K. Lee, and L. A. Symons, BThe face-inversion effect as
   1.142 +    a deficit in the encoding of configural information: Direct
   1.143 +    evidence,[ Perception, vol. 29, no. 2, pp. 159–170, 2000.
   1.144 +  - M. B. Lewis, BThatcher’s children: Development and the Thatcher
   1.145 +    illusion,[Perception, vol. 32, pp. 1415–21, 2003.
   1.146 +  - E. McKone and N. Kanwisher, BDoes the human brain process objects
   1.147 +    of expertise like faces? A review of the evidence,[ in From Monkey
   1.148 +    Brain to Human Brain, S. Dehaene, J. R. Duhamel, M. Hauser, and
   1.149 +    G. Rizzolatti, Eds. Cambridge, MA: MIT Press, 2005.
   1.150 +
   1.151 +
   1.152 +
   1.153 +
   1.154 +heee~eeyyyy kids, time to get eagle'd!!!!
   1.155 +
   1.156 +
   1.157 +
   1.158 +
   1.159 +
   1.160 +* Ullman 
   1.161 +
   1.162 +Actual code reuse!
   1.163 +
   1.164 +precision = fraction of retrieved instances that are relevant
   1.165 +  (true-postives/(true-positives+false-positives))
   1.166 +
   1.167 +recall    =  fraction of relevant instances that are retrieved
   1.168 +  (true-positives/total-in-class)
   1.169 +
   1.170 +cross-validation = train the model on two different sets to prevent
   1.171 +overfitting. 
   1.172 +
   1.173 +
   1.174 +
   1.175 +
   1.176 +
   1.177 +** Getting around the dumb "fixed training set" methods
   1.178 +
   1.179 +*** 2006 Learning to classify by ongoing feature selection
   1.180 +    
   1.181 +    Brings in the most informative features of a class, based on
   1.182 +    mutual information between that feature and all the examples
   1.183 +    encountered so far. To bound the running time, he uses only a
   1.184 +    fixed number of the most recent examples. He uses a replacement
   1.185 +    strategy to tell whether a new feature is better than one of the
   1.186 +    corrent features.
   1.187 +
   1.188 +*** 2009 Learning model complexity in an online environment
   1.189 +    
   1.190 +    Sort of like the heirichal baysean models of Tennanbaum, this
   1.191 +    system makes the model more and more complicated as it gets more
   1.192 +    and more training data. It does this by using two systems in
   1.193 +    parallell and then whenever the more complex one seems to be
   1.194 +    needed by the data, the less complex one is thrown out, and an
   1.195 +    even more complex model is initialized in its place.
   1.196 +
   1.197 +    He uses a SVM with polynominal kernels of varying complexity. He
   1.198 +    gets good perfoemance on a handwriting classfication using a large
   1.199 +    range of training samples, since his model changes complexity
   1.200 +    depending on the number of training samples. The simpler models do
   1.201 +    better with few training points, and the more complex ones do
   1.202 +    better with many training points.
   1.203 +
   1.204 +    The more complex models must be able to be initialized efficiently
   1.205 +    from the less complex models which they replace!
   1.206 +
   1.207 +
   1.208 +** Non Parametric Models
   1.209 +
   1.210 +*** Visual features of intermediate complexity and their use in classification
   1.211 +
   1.212 +*** The chains model for detecting parts by their context
   1.213 +
   1.214 +    Like the constelation method for rigid objects, but extended to
   1.215 +    non-rigid objects as well.
   1.216 +
   1.217 +    Allows you to build a hand detector from a face detector. This is
   1.218 +    usefull because hands might be only a few pixels, and very
   1.219 +    ambiguous in an image, but if you are expecting them at the end of
   1.220 +    an arm, then they become easier to find.
   1.221 +
   1.222 +    
   1.223 \ No newline at end of file