I. Installation In order to use the code for Joint Categorization and segmentation, you have to install the following toolboxes: _ blocks (0.1.1 or above) _ vlfeat (0.9.18 or above) _ maxflow-v3.01 or above (Miki Rubinstein Wrapper available on Matlab FileExchange) _ graphAnalysisToolbox-1.0 or above _ GCMex wrapper for maxflow. NB : Don't forget to add them in Matlab's search path
(0) Everything is parametrized in a JCaS() object. You can create one and explore the structure and the options available.
a.Image/ground truth format : _ Put the images in a directory _ Put the ground truth labelings in another one, with .mat format being a matlab array of integers from 1 to the number of classes, and if it exists 0 as the void class.
b. Adding the database: _ Add your database ine the @jcas/makedb.m file and
c. Parameters _ All parameters can be modified in the Initialization.m file. To run the code without any further modification, just run the script. _ You can transparently change the parameters, and the code will take care of reusing what was previously computed to save computation time/storage space.
d. Force recomputation If you want to recompute some part of the algorithm,
e. Further modification Most of the options are in a single file. You can add your own superpixels/unary/topdown features in the compute*.m files in @jcas dir.
III. Modes You can change the jcas.mode option to the following : 0 = Unary only 1 = Unary and pairwise 2 = Unary + pairwise + linear topdown from ECCV12 paper 3 = Unary + pairwise + linear topdown from ECCV + label cost 4 = Unary + pairwise + linear topdown + label cost (= topdown histogram norm) 5 = Unary + pairwise + label cost only 6 = Unary + pairwise + intersection kernel (PAMI) 7 = unary + pairwise + linear topdown + Unary on words (ECCV) 8 = unary + pairwise + linear topdown + CRF on words (ECCV) (Under construction)