LFW dataset:
We randomly select 2000 images from LFW dataset . Out of these, 1000 images are used for creating training pairs and the remaining (unseen) 1000 for testing pairs. The annotations are collected for 10 attributes, with 500 training and testing pairs per attribute. In order
to minimize the chances of inconsistency in the dataset, each image pair is got annotated from 5 trained annotators, and final annotation is decided based on majority voting.

Training and Testing pairs for each attribute are given.
The images are named 1.jpg ..2000.jpg

for labeltest as well as labeltrain
we compare image1 with image2.


image_compare:
First dimension:  comparisionID.
Second dimension: name of image1 which was used in comparision.
Third dimension:  name of image2 which was used in comparision.

attribute_strengths:
First dimension :  comparisionID.
Second dimension:  number of annotators out of 5 who said image1 has more strength of attribute than image2.
Third dimension:   number of annotators out of 5 who said image1 has less strength of attribute than image2.
Fourth dimension:  number of annotators out of 5 who said image1 has equal strength of attribute as image2.

If you use this dataset please cite: 

@InProceedings{SandeepCVPR2014,
	title = "Relative Parts: Distinctive Parts for Learning Relative Attributes",
	author = "Ramachandruni N.Sandeep,Yashaswi Verma,C.V.Jawahar",
	booktitle = {IEEE Conference on Computer Vision and Pattern Recognition (CVPR)},
	year      = {2014}
}


