###################################
FEATURES FOR MIR-FLICKR 1M DATASET
Nitish Srivastava, University of Toronto.
##################################

The following features were extracted for the images:
(1) Color Layout Descriptor (192 dimensions)
(2) Color Structure Descriptor (256 dimensions)
(3) Scalable Color Descriptor (256 dimensions)
(4) Homogenous Texture Descriptor (43 dimensions)
(5) Edge Histogram Descriptor(150 dimensions)
(6) Gist (960 dimensions)
(7) Pyramind Histrogram of Words (using SIFT) (2000 dimensions)

The image dataset consists of 3857-dim vectors where these features have been
concatenated in the order shown above.

The text features are word counts of 2000-most frequent tags.

These features were used to obtain the results reported in :
Nitish Srivastava and Ruslan R. Salakhutdinov, "Multimodal Learning with Deep
Boltzmann Machines", Neural Information Processing Systems (NIPS 2012).

If you use these features in your experiments, please cite this paper.