Joint Calibration v1.0
================================================================================

Davide Modolo, Alexander Vezhnevets, Olga Russakovsky, Vittorio Ferrari
d.modolo@sms.ed.ac.uk

This software was developed under 64-bit Linux with Matlab R2015b.There
is no guarantee it will run on other operating systems or Matlab versions
(though it probably will).

If you use this software for academic research, please cite [1].

All software are released as free software for non-commercial applications 
under the terms of the GNU General Public License version 3 (GPLv3), see 
<http://www.gnu.org/licenses/>, which is also noted in each individual file. 
For commercial applications, contact the author for licensing options.

This software doesn't need any external packages.
    
If you find any bugs or have any comments, please e-mail D. Modolo



Introduction
--------------------------------------------------------------------------------
This release implements the joint calibration technique presented in [1]. 
The algorithm requires Exemplar-SVMs window scores and labels and outputs
the set of thresholds, one for each Exemplar-SVM, that minimize the 
optimization problem of [1], eq. 2.

To quickly test our method we have included a demo file with an example and 
the corresponding solution. 


Quick start
--------------------------------------------------------------------------------
Let us assume that <DIR> is the directory where you uncompressed the release
archive. In the Matlab prompt, type:

    cd <DIR>
    mex CXXFLAGS='$CXXFLAGS -std=c++11 -fPIC' fast_unique.cpp
    demo

Note: The release archive already contains a precompiled mex file for 64-bit
Linux, so you only need to compile it again if you are on another platform 
(e.g. 32-bit Linux or Windows).

The demo.m script finds the optimal thresholds for a small problem of 6
Exemplar-SVMs and 16 windows (6 positive and 10 negative). 

joint_calibration.m is the main function of this release and the only one that 
should be used. This function operates in two phases. First, it pre-computes 
lots of data and second, it search for the optimal configuration of thresholds. 
The pre-computation is very important and it is essential to speed up the 
search. Note that this pre-computation can take hours on large problems
of few thousands Exemplar-SVMs and positive windows. 


Changelog
--------------------------------------------------------------------------------

v1.0: First public release



References
--------------------------------------------------------------------------------
[1] Modolo, D. and Vezhnevets, A. and Russakovsky, O. and Ferrari, V.
    Joint calibration of Ensemble of Exemplar SVMs,
    CVPR 2015, Boston, USA