Update on 01/20/2014: A small bug in ranksvm_with_sim.m has been fixed. The predictors below have been re-trained using the updated code. See documentation in ranksvm_with_sim.m for more details. We provide the two datasets (osr and pubfig), the learnt relative attributes and their predictions used in the following paper: Devi Parikh and Kristen Grauman Relative Attributes International Conference on Computer Vision (ICCV), 2011 (Oral). @InProceedings{relative_attributes, author = {D. Parikh and K. Grauman}, title = {Relative Attributes}, booktitle = {IEEE International Conference on Computer Vision (ICCV)}, month = {Nov}, year = {2011} } If you use any part of this data, please cite the above paper. Each directory (osr and pubfig) contain the following: images: The images for the pubfig dataset contain cropped faces re-sized to be 256x256. These faces have been extracted from a subset of the images in the PubFig dataset (http://www.cs.columbia.edu/CAVE/databases/pubfig/) of Kumar et al. The images for the osr dataset are directly taken from the outdoor scene recognition dataset of Oliva and Torralba (http://people.csail.mit.edu/torralba/code/spatialenvelope/). A link to a downloadable folder of images has been provided in images.txt data.mat: Loading this file in MATLAB creates the following variables in the workspace: im_names: Names of the images attribute_names: Names of the attributes class_names: Names of the categories class_labels: The ground-truth category label of each image feat: Features extracted for each image (gist for osr, gist concatenated with color for pubfig) binary_predicates: A binary matrix indicating the presence of an attribute for each category. This was used as ground truth to train binary attribute predictors. relative_ordering: The relative ordering of all categories for each of the attributes. Low values indicate a weak relative strength of the attribute in the class. Equal values indicate similar relative strengths of the attribute. Let's consider attribute i and classes j and k. relative_ordering(i,j)>relative(i,k) indicates that class j has a stronger presence of attribute i than class k. This was used as ground-truth to train relative attribute predictors using a learning to rank formulation as described in the paper. relative_att_predictors: The weight matrix of learnt linear relative attribute predictors. relative_att_predictions: The score of each image for each relative attribute. relative_att_predictions = feat*relative_att_predictors. used_for_training: A binary indicator that specifies which of the images were used to train the relative attribute predictors.