README

SUN Attribute Database Classifiers v2

This release includes pre-computed classifiers for all of the attributes of the SUN Attribute dataset as described in the paper:
Genevieve Patterson, James Hays. SUN Attribute Database: Discovering, Annotating, and Recognizing Scene Attributes. Proceedings of CVPR 2012. 

How to Use:

A vector of attribute confidences for a given query image can be computed using the following Matlab function - 
	confidence_vector = classify_attribute(test_image, image_kernel_save_path, image_path, image_feat_path, classifier_path)
Instructions for how to use this function can be found in the header of classify_attribute.m. 

	IMPORTANT: set up your operating paths in attributes_globals.m before trying to classify any test images.
			Add the path to all of the SUN dataset source code and Attribute dataset source code to the working path
			before trying to classify attributes. Also setup the vlfeat library by running the command:

			vl_setup('noprefix');

	Known issue: If you receive an error when the geometric color histogram features are being calculated that looks like this:
			"/bin/bash: ../segment_pedro/segment: No such file or directory"
			Go to the file "SUN_source_code_v2/code/GeometricContext_dhoiem/src/im2superpixels.m"
			and change line 8 to -- "segcmd = 'segment_pedro/segment 0.8 100 100';"

	

Required Packages (these should be included in Matlab's working directory):
	SUN Database source code v2
	http://people.csail.mit.edu/jxiao/SUN/source_code/

	Primal SVM
	http://olivier.chapelle.cc/primal/


How to re-compute classifiers:

To generate new test/train sets run:
	generate_train_test_set.m

To calculate features for all of the SUN Attribute DB images and create the required feature matricies run:
	calc_feature_set.m

To calculate the required kernels run:
	calc_master_kernel.m
	calc_master_combined_kernel.m

To train the attribute classifiers, run the following function for each attribute:
	train_attribute_svm.m


-- 
Genevieve Patterson, 12/18/2012