Natural Scene Images Dataset with Inconsistent Rankings

The original pictures can be downloaded in http://www.lamda.nju.edu.cn/data_MIMLimage.ashx .

This dataset includes inconsistent rankings of 2000 natural scene images. There are nine possible labels associated with these images, i.e., plant, sky, cloud, snow, building, desert, mountain, water and sun.
Ten human rankers are requested to label the images. For each image, they first select from nine candidate labels what they think are relevant to the image, and then rank the relevant labels in descending order of relevance to the image. Each human ranker makes his/her decisions independently, i.e., each ranker can not see the results from other rankers.

In "pic_scene.mat", pic is 1*n struct where n is the number of instances, in natural scene dataset n is 2000.

pic(i) is the inconsistent rankings of i-th image, e.g., "pic(1)" refers to "1.jpg" in origin dataset.

Each of pic(i).label is a matrix with size r*c, where r is the number of rankers and c is the number of labels (including the virtual label).
In natural scene dataset r is 10 and c is 10.
The labels is:
1	plant
2	sky
3	cloud
4	snow
5	building
6	desert
7	mountain
8	water
9	sun
10	virtual

The pic(i).label(j, :) is the ranking labeled by j-th ranker in descending order of relevance to the image.
The labels ranked before virtual label are relevant labels. And those ranked after virtual label are irrelevant labels, with tie ranks.
Such as pic(1).label(1, :), which is the ranking labeled by the first ranker for the first image:
6	2	10	1	3	4	5	7	8	9
In this ranking, labels 6(desert) and 2(sky) are relevant labels, while 1(plant), 3(cloud), 4(snow), 5(building), 7(mountain), 8(water) and 9(sun) are irrelevant labels.
And 6(desert) > 2(sky).

