INFO:
This is the extracted features of the paper
T. Matsukawa, T. Okabe, E. Suzuki, Y. Sato, 
"Hierarchical Gaussian Descriptor for Person Re-Identification", 
In Proceedings of IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2016.
 
Usage
=====
Each .mat file contains the feature_all size: [n_all dim] where n_all is the number of images 
and dim is the dimension of GOG features. 
(The order of the images is same as the allimagesnames in the DB). 
 
Note that the feature vectors are not normalized (because the normalization depens on training data). 
To reproduce the results of the paper, 
you need to normalize the feature vector (Eq.6 in the paper). 
For example, 

Divide the feature_all into training dataset (Xtrain size. [n_train dim]) and test dataset (Xtest size.[n_test dim])
where n_train/n_test are the number of training/test images. 

meanX = mean(Xtrain, 1); % mean vector 

% X -- both Xtrain and Xtest
% Y -- output. 
 
Y = ( X - repmat(meanX, size(X,1), 1)); % Mean removal
for dnum = 1:size(X, 1)
    Y(dnum,:) = Y(dnum, :)./norm(Y(dnum, :), 2); % L2 norm normalization
end

DB
==
'./DB' contains dataset information.  
    - allimagenames -- all image names of the dataset
    - traininds_set/testinds_set -- index of the training/test images for each division
                                    (index order is the same as allimagenames)
    - trainimagenames_set/testimagenames_set -- image names of the each traing/test division
    - trainlabels_set/testlabels_set -- person IDs of the trainig/test images for each division
    - traincamIDs_set/testcamIDs_set -- camera IDs of the trainng/test images for each division
