(c) 2010 - Emmanuel Goossaert
Under GNU License 3.0

--- Get started ---

The source code for the Lambertain Color Model that I implemented is available under the GNU
License 3.0.
It can be browsed here: http://github.com/goossaert/computer-vision/tree/master/lambertain/,
and downloaded here: http://github.com/goossaert/computer-vision/archives/master.

I have run experiments with this model, that I have presented in a post on my blog:
http://codecapsule.com/2010/09/01/background-segmentation-lambertain-color-model-opencv/

After you have downloaded the source code, unpack it and open the Makefile to set the library
and include locations to the directory containing OpenCV 2.1 on your computer:

$ tar zxvf lambertain.tgz
$ cd lambertain
$ vim Makefile
$ # ... update the library and include locations to match the one on local machine

Now just run make to compile:

$ make

This will create three executables:

* background. Performs background segmentation. You can run the program without option to get the
  list of parameters it requires. Also, the shellscript “test_background.sh” runs the program on
  the dataset from the article.
* color. Performs color segmentation. You can run the program without option to get the list of
  parameters it requires. Also, the shellscript “test_color.sh” runs the program on the dataset
  from the article.
* videocapture. A small utility that allows you to capture images with a webcam. This is what I
  used to create my training set, so if you have a webcam, you can use this utility to create
  your own dataset and rapidly try the algorithm. If you do not have a webcam, I have included the
  training images that I used with the source code, so that you can experiment with that.

If you are working on Windows, you will need to create a project in the IDE that you use, and add
the model files to it.


--- Possible improvements and optimizations ---

* Threshold selection. Right now, the algorithm to select the thresholds is very inefficient.
  Indeed, I simply sort the data and use indexes to find the thresholds, yielding to a
  complexity of a O(n lg n). This could be optimized by using a selection algorithm, which
  would bring complexity down to O(n).
* Multiply, not divide. Multiplying by the inverse of the denominators instead of dividing
  by these denominators would speed up computations, as explained in Section 7 of
  Horprasert et al. (1999).
* Global variance. Using global average variance instead of local variance would speed up
  computations, as explained in Section 7 of Horprasert et al. (1999).
* Clustering Detection Elimination. I have not implemented the technique covered in
  Section 5 of Horprasert et al. (1999), which rectifies the erroneous chromaticity
  distortion values. This would allow the model to produce more accurate results.


--- References ---

The scientific publications used to implement this model can be found here:

T. Horprasert, D. Harwood, and L.S. Davis. A statistical approach for real-time robust
background subtraction and shadow detection. IEEE ICCV, 1999.
http://citeseerx.ist.psu.edu/viewdoc/download?doi=10.1.1.87.1244&rep=rep1&type=pdf

Y. Yacoob and L.S. Davis. Detection and Analysis of Hair. IEEE Transactions on Pattern
Analysis and Machine Intelligence, 1164-1169, 2006.
http://www.umiacs.umd.edu/~yaser/pamis.pdf