scene_categories_spatial_rationale_annotations.zip is a zip file containing MATLAB annotation data from our experiment on the Fifteen Scene Categories dataset [1], which can be downloaded from:
	http://www-cvr.ai.uiuc.edu/ponce_grp/data/scene_categories/scene_categories.zip

The directories in this archive correspond to the fifteen scene categories in the dataset.  Each directory contains all annotations that were LABELED (i.e., possibly mislabeled) with that category name.

An example of a MATLAB data file in this archive is insidecity/1727.mat.  If this file is loaded into MATLAB, a variable called "annotation" is created:
	
	>> load('insidecity/1727.mat');
	>> annotation
	
	annotation =
		
		filename: 'street/image_0209.jpg'
		label: 'insidecity'
		polygonsX: {[11x1 double]  [7x1 double]}
		polygonsY: {[11x1 double]  [7x1 double]}

The "filename" attribute is the image from the dataset [1] that the annotator was given to annotate (in this case, it was an image from the category 'street', meaning this image was mislabeled according to the original labels).

The "label" attribute is the label the annotator gave.  All files in the "insidecity" directory will have the label "insidecity", as the files are put into directories according to annotator label.

Attributes "polygonsX" and "polygonsY" contain the image coordinates of the vertices of the annotator's rationale polygon(s).  In this case, the annotator has marked two polygons as rationales; the first having 11 vertices and the second having 7 vertices.


[1] L. Fei-Fei and P. Perona. A Bayesian Hierarchical Model for Learning Natural Scene Categories. In CVPR, 2005.