Stereo Package


The Stereo Package is released for academic research only under the Lesser General Public License.  This package is a free software as described in the LGPL license.  For purposes other then academic research please contact  Talya Meltzer.

The Package

This package was used to produce the results presented in "Globally Optimal Solutions for Energy Minimization in Stereo Vision using Reweighted Belief Propagation".
It includes:
1. Marshall Tappen's code in c++ for running BP on a 2-dimensional grid: We applied few modifications to make it run TRBP.
2. Matlab code: implemented by Chen Yanover and Talya Meltzer, used to analyze the TRBP results and extract the minimal energy configuration for the standard bench mark images
3. The bench mark image sequences

Download

The stereo package: stereo.tgz

Our Matlab code requires also installation of:
1. Chen Yanover's Matlab sparse cell class
2. Kevin Murphy's Bayes Net Toolbox for Matlab

Installation

The package includes the following directories:
1. StereoMatch: Marshall's c++ code with our BP to TRBP modifications
2. imagedirs: The benchmark image sequences and some of Marshall's input files for his code
3. imagedirs/Scripts: Our automatically created (using our matlab wrapping) input files for Marshall's code + Marshall's input files for running other optimization methods
4. mFiles, mexFiles: Our matlab interface. Allows to run Marshall's code for any image sequence and any energy parameters, and analyzes the results in order to extract the global optimum.
5. shScripts: The shell-scripts we use from our matlab wrapping to run Marshall's code.

After untarring the package under your directory, edit the following variables in the file mFiles/stereoMAP.m:
SPARSE_CELL_DIR - the path where Chen's sparse cell class is installed
BNT_DIR - the path where Kevin's BNT package is installed
ROOT_DIR - the path where the stereo package is installed

Usage

Extracting the global optimum for a certain sequence and parameters:

In the matlab-script mFiles/stereoMAP.m define and comment out the image parameters: seq_name and numStates, and the energy parameters s,T,P.
Possible values for image parameters:

seq_namenumStates
map30
tsukuba16
venus20
sawtooth20

"numStates" is the number of possible disparities for each pixel in the image. It is defined for each sequence in the files imagedirs/<seq_name>/param_in.txt, and can be changed there.
The energy parameters we used are presented in the paper with the minimal energies extracted.

Reproducing the results presented in the paper:

The matlab-script mFiles/allMinEConfs.m runs the script mFiles/stereoMAP.m for all the paramters presented in the paper.
Make sure those parameters (seq_name, numStates, s, T, P) are commented out in stereoMAP.m !


Questions, suggestions, comments and bug report - please send to  Talya Meltzer

Back to Talya Meltzer's home page