Weakly supervised graph based semantic segmentation by learning communities of image-parts

For the initial segmentation step, use the JSEG segmentation. Available at http://old.vision.ece.ucsb.edu/segmentation/jseg/software/

• phow-bow.m : Uses VLFEAT to compute the PHOW - bag of words features from the dataset. Most of the code is taken from phow_caltech101.m example of VLFEAT.

  Uses -> standarizeImage.m: Resizes the images -> getImageDescriptor.m: Computes the descriptor histograms from the input image

• regional_phow.m: Uses VLFEAT to compute the features on the segmented regions. Requires the images to be segmented beforehand. Should be used on training and test data separately.

  Uses -> build_graph_singleLevel_2.m: Creates a graph for a segmented image at a particular level

• all_border.m: Creates the border matrix on segmented images. Should be used on training and test data separately.

• C++ part -> Main folder: Run the main.cpp to read the outputs of previous matlab scripts and create the communities.

   Requires OpenCV, Matlab Engine and NCUT codes. NCUT available at http://www.timotheecour.com/software/ncut/ncut.html
   	Uses Matlab engine to call Ncut method. Uses Ncut_C.m to call the method.
   	Arguments: 
   	argv[1] : N1, total number of regions in training set.  argv[2] : N2, total number of regions in test set.
   	argv[3]: M, number of words.
   	argv[4]: DB1, number of images in training set.
   	argv[5]: num_class, number of classes in 
   	argv[6]: L, class similarity parameter
   	argv[7]: seg_level, argv[8]: DB_name , argv[9]: Feature -> Address manipulation strings
   	argv[10]: max_region1, maximum number of regions in one image in training set.
   	argv[11]: N_conn,  weighted network parameter, number of connected nodes in the network
   	argv[12]: num_comm, number of communities
   	argv[13]: knn_consider
   	argv[14]: DB2, number of images in testing set.
   	argv[15]: max_region2, maximum number of regions in one image in testing set.
  

• Define_Y_initialization_community_labels.m : to find alpha_l, use the method -> linear_search_alpha.m

• transfer_label_to_test_images.m: Assigns labels to the detected communities on test images

Extras

solve_opt.m
CommRelation.m
number_of_regions.m
comm_driven_graph.m