## This file explains the synthetic dataset used for learning
## parameters in discriminative fields 

DATASET DESCRIPTION

The dataset consists of 200 noisy images (50 noisy versions of each of
the four binary base images). Two types of noise were used:

(A) Gaussian noise: Each true image was corrupted with a gaussian
noise. Details of the parameters are given in the following paper:

S. Kumar and M. Hebert, "Discriminative Fields for Modeling Spatial
Dependencies in Natural Images," in proc.  advances in Neural
Information Processing Systems (NIPS), December 2003

Four matlab files containing 50 noisy images each and the corresponding
ground truth label file (in parenthesis) are:

gaussNoise/dataMat_multSyn1N90.mat  (totalLabels_synIm1.mat)
gaussNoise/dataMat_multSyn2N90.mat  (totalLabels_synIm2.mat)
gaussNoise/dataMat_multSyn3N90.mat  (totalLabels_synIm3.mat)
gaussNoise/dataMat_multSyn4N90.mat  (totalLabels_synIm4.mat)


(B) Bimodal noise: Each base image was corrupted by a mixture of two
gaussians noise. Details of the parameters are given in the above
mentioned paper.

Similar to the above case, four matlab files containing 50 noisy
images each and the corresponding ground truth label file (in
parenthesis) are:

biModalNoise/dataMat_multBimodalSyn1.mat (totalLabels_synIm1.mat)
biModalNoise/dataMat_multBimodalSyn2.mat (totalLabels_synIm2.mat)
biModalNoise/dataMat_multBimodalSyn3.mat (totalLabels_synIm3.mat
biModalNoise/dataMat_multBimodalSyn4.mat (totalLabels_synIm4.mat)

NOTES: 
1. The data has already been normalized to have zero mean and variance
one using the training set mean and variance.

2. Each base image is a binary image and the labels are in format {1, -1}.

3. For training, we used 10 noisy versions of base image 1 for both
the noises and the testing was done on the rest of the images.





