0ddoes / multilevel-object-closure-detection-by-superpixels Goto Github PK

General info:
-------------

This is the code for the CSC2523 course project, by Hanwen LIANG and Han Meng, winter 2018 at University of Toronto. We intend to use the work of “Optimal Contour Closure by Superpixel Grouping", Alex Levinshtein, Cristian Sminchisescu, Sven Dickinson. ECCV 2010” as the starting point to design an recursive algorithm to detect all possible objects and their closures for an image. Now the tree structure has been completed.

We offer sincere thanks for Professor Sven Dickinson and Alex Levinshtein for their guidance and help in this project!

The code was based on most from the original code from “Optimal Contour Closure by Superpixel Grouping", Alex Levinshtein, Cristian Sminchisescu, Sven Dickinson. ECCV 2010”. We added some changes and the recursive object-closure detection part.

When using this code please reference the above paper.

Our implementation included code from other sources:
1) Pb and globalPb from University of Berkeley
2) Normalized cuts
3) Contour fragment code from Andrew Stein
4) Parametric Maxflow from Vladimir Kolmogorov et al
5) Matlab helper functions from Peter Kovesi
6) MatlabBGL library for graph operations from David Gleich
7) SLIC super pixels from VLFeat

We acknowledge the authors of the above software.
Our code, as well as the code included above is covered under GPL (see LICENSE)

Before running the program on MATLAB, you will need to have the proper C++ compilers, MATLAB software, and VLFeat open source package installed.

Quick startup guide:
--------------------

1) Start the VLFeat library.
1) Run the command 'DefinePath' from the Closure directory in Matlab
2) Run make.m (if you haven't done so before on this system)
3) Use the function 'ClosureMain' to compute closure (see header for usage)

The code was tested on Matlab 7.8 on Linux 32bit, Linux 64bit, and Windows.

Note:

1) For computational efficiency, we save the preprocessing results of the various stages (Pb, globalPb, superpixels). 
However, this can cause unexpexted behaviour. For example, if superpixel were computed with Pb, but then one desides to use 
globalPb, the superpixels won't be recomputed since the old superpixels are already stored.

2) The by default the code uses Pb and not globalPb to compute contours.
globalPb can be used by setting the use_gpb = true flag. We provide a compiled version of globalPb for Matlab 7.8
for 32 and 64 bit Linux. For a different platform the user will need to compile the globalPb code manually and put it under the globalPb directory. The code can be downloaded from : http://www.eecs.berkeley.edu/Research/Projects/CS/vision/grouping/

3) For the mainfunction ./Multilevel-Object-Closure-Detection-By-Superpixels/ClosureMain.m, we generate name-specific file automatically to store related files. And we use "XXX_num1_num2_num3_multisolutions.jpg"(num1.2.3) to represent the hierarchy of images.
 
Example:
--------
Run the following from matlab after completing the setup steps above:

>> ClosureMain('airplane.png', '.', 100, 10, 0.05, 'turbo');

This will run closure detection on the image 'airplane.png' in the current directory and store the results in the
current directory. This example uses 100 superpixels, returns at most 10 solutions, and uses an edge threshold (T_e)
of 0.05. The turbopixels method will be used to generate superpixels.

This will generate airplane_solution_???.jpg files, where each one contains a binary mask for one closure hypothesis.
The solutions are ordered by increasing closure cost. We also generate airplane_multiplesolutions.jpg file the visualizes all the solutions in one image.