------------------------------------------------------------------
Readme file for image data and learned distributions
------------------------------------------------------------------


IMAGES AND CONFIGURATIONS

The training images are located in ImageData. Each image has a corresponding 
manually determined edge configuration, in ManualEdgeData.
The configuration is in MAT format. When loaded into a Matlab program (for a 
file called <name>.mat, do "load <name>"), a matrix named edges is obtained. 

The first dimension (row) corresponds to limb number: 
    (1=torso, 2=thigh 3=calf, 4=upper arm, 5=lower arm, 6=head)
For each row, the four columns (second dimension) corresponds to the four 
    corners of the limb (endpoints of two edges).
Each corner has an x and y location in the image (third dimension):
    (1=x location, 2=y location)


LEARNED DISTRIBUTIONS

In Distributions, mat files containing the learned distributions can be 
found. When loading edge.mat or ridge.mat, a matrix RATIO is obtained. 
The first dimension is the filter response. The filter response is a 
    real-valued number, and the bin corresponding to filter response x is 
    found as:
         EDGE_NUNITS*EDGE_NBINS +  floor(x*EDGE_NBINS) + 1
    or, in the case of ridge:
         RIDGE_NUNITS*RIDGE_NBINS +  floor(x*RIDGE_NBINS) + 1
    The constants are:
         EDGE_NBINS = 40;      % number of bins per unit edge contrast      
         EDGE_NUNITS = 2;      % number of units edge contrast covered
         RIDGE_NBINS = 20;     % number of bins per unit ridge contrast
         RIDGE_NUNITS = 4;     % number of units ridge contrast covered
The second dimension is the image level. Since all levels have the same 
    distribution, this dimension is of length 1.
The third dimension is the limb number:
    (1=torso, 2=thigh 3=calf, 4=upper arm, 5=lower arm, 6=head)

The file flow.mat contains the two matrices BG and FG (since ratio cannot be 
computed, see Section 3.3 in [2]).
The first dimension is filter response as with ridge and edge. It is obtained
    in the same way: 
         FLOW_NUNITS*FLOW_NBINS +  floor(x*FLOW_NBINS) + 1
    where
         FLOW_NBINS = 0.2;     % number of bins per unit flow difference
         FLOW_NUNITS = 255;    % number of units flow difference covered
The second dimension is image level:
    (1=level 0, 2=level 1, 3=level 2, 4=level 3)
The third dimension  is the limb number:
    (1=torso, 2=thigh 3=calf, 4=upper arm, 5=lower arm, 6=head)
    Of course, BG has only one distribution, and the length of the third 
    dimension is thus 1.


RELATED PUBLICATIONS

[1] Hedvig Sidenbladh. Probabilistic Tracking and Reconstruction of 3D Human 
    Motion in Monocular Video Sequences. Doctoral Thesis TRITA-NA-0114, 
    ISBN 91-7283-169-3, Dept. of Numerical Analysis and Computer Science, 
    KTH, Stockholm, Sweden 2001.

[2] Hedvig Sidenbladh and Michael J. Black. Learning the statistics of people 
    in images and video. International Journal of Computer Vision (IJCV), to 
    appear. (Version as of April 2001 available as TRITA-NA-P0118, Dept. of 
    Numerical Analysis and Computer Science, KTH, Stockholm, Sweden 2001.) 

[3] Hedvig Sidenbladh and Michael J. Black. Learning image statistics for 
    Bayesian tracking. In IEEE International Conference on Computer Vision 
    (ICCV), vol 2, pp 709-716, Vancouver, Canada 2001. 

