This software package detects shadows from single images by using an approach described in the following paper:

J.-F. Lalonde, A. A. Efros, and S. G. Narasimhan, "Detecting ground shadows in outdoor consumer photographs," presented at the European Conference on Computer Vision, 2010.

Please cite this paper if you use this code in your work.

Getting started

1. Install the required packages (see below), and compile them;

2. From the mycode directory, run:

   $ setPath
   $ demoShadowDetection
   

3. The results should appear as they are computed. Enjoy!

Changing the input image

If you want to make it work on your own image, here's what you should do:

1. Copy your image in the img directory (should have .jpg extension);

2. Change the imgName variable with the new filename (leave the .jpg extension out);

3. If you want to detect shadows on the ground only:

   3.1 Compute the ground probability map (e.g. with Geometric Context, see below); 3.2 Save the ground probability map as variable groundProb in a .mat file; 3.1 Copy the .mat file containing the variable groundProb in the img directory; 3.2 Rename the file to <image name as in step 2>-groundProb.mat.

4. Watch the output while the program is running to make sure there are no errors or warnings;

5. That's it!

Requirements

Required 3rd-party software not included in this package

• My "utils" package, available at: https://github.com/jflalonde/utils/.
• Lightspeed matlab toolbox by Tom Minka, need at least version 2.4.
• Graph cut wrapper by Shai Bagon which uses C++ code by Olga Veksler.

If you put these 3rd-party packages in a pathUtils folder in the same base folder as the shadowDetection, they should be picked up automatically by setPath.

Make sure you follow the respective instructions to install those packages.

Required 3rd-party software included in this package

In addition, this code uses the following freely-available matlab code:

• Bilateral filtering by Douglas R. Lanman
• Boundary extraction from Andrew Stein
• Fast nearest-neighbor search by Luigi Giaccari, also available here;
• Boosted decision tree by Derek Hoiem

Compilation

Check within each of the following directories and make sure you compile the .mex files. They are required for this software to run:

• Boosted decision trees

  $ cd 3rd_party/boost
  $ mex treevalc.c
  

• Nearest neighbor

  $ cd 3rd_party/nearestneighbor
  $ mex BruteSearchMex.cpp
  

Optional 3rd-party software not included in this package

• Geometric context by Derek Hoiem.

Tuning the parameters

There are a few parameters you might want to fiddle with in order to tune the results to your liking.

• Resize the image to ~600 pixels along the vertical dimension, since computing features is likely to be prohibitively expensive for images of larger size.

• If the algorithm fails to detect shadows that are too soft, try reducing the image size, or modify the Canny edge detection parameters to make it fire on that edge. We are, after all, looking for shadow edges. You will find the Canny edge detection call at line 39 of applyLocalBoundaryClassifier.m.

• If the detected shadow edges are too discontinuous, you can experiment with the over-segmentation parameters to obtain larger, more continuous segments. You will face a trade-off, however: larger segments will be less precise at capturing small details. For this, look at extractImageBoundaries.m and experiment with the prctMax parameter.

News

• 10/03/2012: Code is now on github, see commit messages for news and changes!
• 05/24/2011: The code should now work without requiring the Matlab Statistics Toolbox (thanks to M. Chen for helping me figure this out!). It's best to re-download the code completely. The following changes have been made:
  • The data/bdt-eccv10.mat file has been changed to use built-in 'structs' instead of classtreereg.
  • The skewness.m function has been added to replace the toolbox's.