Baselines for Image Annotation.
Ameesh Makadia, Vladimir Pavlovic, and Sanjiv Kumar.
International Journal of Computer Vision, Volume 90, Issue 1, October 2010.
Download: (PDF)

Datasets:
ESP Game (espgame.org)
Our evaluation dataset was a subset of the corpus available at (http://www.hunch.net/~jl/).
To replicate our dataset exactly, you will need to download esp_data.mat.
This Matlab data file contains four variables:
'dict': a cell array containing all the words in the dictionary.
'data': a struct with two fields. 'data(i).keywords' is an index vector into 'dict' listing the keywords associated with image i. 'data(i).file' is the filename of image i.
'train': an index vector for the training set. E.g. data(train) gives the keyword indices and filenames for the training set.
'test': same as 'train' but for the test/evaluation set.


IAPR TC 12
Our evaluation dataset was a subset of the corpus available at (http://eureka.vu.edu.au/~grubinger/IAPR/TC12_Benchmark.html).
To replicate our dataset exactly, you will need to download iapr_data.mat.
This Matlab data file follows the same structure as for the ESP set explained above.