This package contains the elements and mid-level representations
learned for the indoor67 dataset with the "visual element discovery
as discriminative mode seeking" algorithm.  This file was created
by Carl Doersch (cdoersch at cs dot cmu dot edu).

The contents are:
final_elements: the set of elements learned.  It's taken from
                the ds.finmodel variable in the code release.  It 
                contains a w vector for the HOG template, the bias b, 
                and a unique id for each element.

imgs_[train,test]: the testing and training image dataset descriptors,
(not in minimal    in the same format as ds.imgs in the code.  
package)

pooledfeats[train,test]: The training and testing pooled feature vectors.
                         They correspond to ds.[test]poolfeats in the code.
(not in minimal          NOTE THAT THESE HAVE NOT BEEN THRESHOLDED!  You'll
package)                 have to add .5 and threshold them at 0 to get the
                         performance reported in the paper.

ds: ds.conf contains the configuration options and whitening data used to 
    learn the elements, as well as the dataset descriptor.  These will be 
    useful if you try to run the detectors.  If you have the code, you 
    should be able to run the detectors by simply calling 
    detectInIm(final_elements,image).