Skip to content


Folders and files

Last commit message
Last commit date

Latest commit



11 Commits

Repository files navigation

Learning Low-order Models for Enforcing High-order Statistics

This software implements the learning approach introduced in:

Patrick Pletscher & Pushmeet Kohli
Learning Low-order Models for Enforcing High-order Statistics
AISTATS, 2012.

The paper can be downloaded from:

If you use the software in your work, then please consider citing the paper as

  author    = {Pletscher, Patrick and Kohli, Pushmeet},
  title     = {Learning Low-order Models for Enforcing High-order Statistics},
  booktitle	= {Proceedings of the Fifteenth International
               Conference on Artificial Intelligence
               and Statistics ({AISTATS}) 2012},
  pages     = {886--894},
  year      = {2012},
  publisher = {JMLR: W\&CP 22},
  address   = {La Palma, Canary Islands},
  editor    = {Neil Lawrence and Mark Girolami}  

1. Installation

The implementation splits the work into three parts: prepare, main and
collect, so that things can easily be run on a cluster for different data
sets and features. Most of the main logic is in grabcut/grabcutMain.m and the scripts
it is calling. If you want to run the main script on a cluster, I recommend
compiling it with something along the lines of the command
(note: you'll need to adapt the mosek path in this file). The step-by-step
guide below describes the workflow assuming you compile the code.

a. obtain mpe_inference and matluster
    cd grabcut
    git clone
    git clone
    cd mpe_inference

b. Install the mosek toolbox in a subfolder grabcut/mosek. You need the "free academic
   license". I would recommend just creating a symlink to a central place
   where you save the mosek toolbox, such as ~/mosek.

c. Obtain the images and corresponding segmentations from:

   Unpack the files into a grabcut/data/ folder. The folder should now contain
   three subfolders: images, images-gt and images-labels.

   Furthermore, install the gsc-1.2 software by downloading it from:

   and put it into grabcut/gsc-1.2. Run


   from within Matlab.

d. In matlab go to preprocess and run
   mex mex_setupTransductionGraphNew.cpp

e. In Matlab run the command:

f. In Matlab run the command:

g. On the command-line compile the main script:

h. From the command-line submit the jobs to the cluster:

i. Once all the jobs have successfully completed, run the collection from
   within Matlab:

2. Authors

The software was written by:
Patrick Pletscher and Pushmeet Kohli

We would like to acknowledge:
- Sebastian Nowozin for input about Mosek and pruning strategies.
- Aditya Khosla for general feedback.

For convenience we included the gsc-1.2 package, which can be obtained from

If you use the gsc code, please also acknowledge the relevant paper:
Geodesic Star Convexity for Interactive Image Segmentation 
V. Gulshan, C. Rother, A. Criminisi, A. Blake and A. Zisserman at CVPR 2010


Label-Count Loss Augmented Inference






No releases published


No packages published