Using color compatibility for assessing image realism
This software package implements the algorithms presented in our ICCV 2007 paper:
J.-F. Lalonde and A. A. Efros. Using color compatibility for assessing image realism. In IEEE International Conference on Computer Vision, October 2007.
It implements both the realism assessment (Sec. 3-4-5) and recoloring (Sec. 6) algorithms.
Please note that most of the code included herein is not actually used anywhere, and was developed as part of the research. I'm including it for completeness. Please refer to the demo/demoColorCompatibility.m file as a starting point to understand what is being called, and how.
- Download 'utils' package and VLfeat (see below);
- Download the compressed database information and extract into demo/data (see below);
- Edit 'getPathName' and modify path to match your local machine.
- Run 'setPath', make sure there are no errors or warnings.
- Run 'demoColorCompatibility'.
Download the compressed database information at
Extract the .zip file (which should contain a 'db' directory) into the demo/data directory. The demo code expects it there.
The 'db' directory contains the following information:
- concatHisto (and concatHistoTextons): concatenanted color and texton histograms for each image in the database. Used to compute the global realism measure.
- objectDb: large struct array containing information for each object/background pair in the database. Used to retrieve the image corresponding to each pair.
Make sure to compile (mex):
BruteSearchMex: cd 3rd_party/nearestneighbor; mex BruteSearchMex.cpp;
EMD: cd 3rd_party/emd; build_emd;
Don't forget to download the "utils" code, available at:
Also, the re-coloring code requires VLFeat, available at:
Version 0.9.14 was used to test this code.
The Berkeley Segmentation code, included within this package, originated from:
All credits go to the original authors of this work.
Right now, the image database is accessible through the webserver: http://balaton.graphics.cs.cmu.edu/jlalonde/colorStatistics/Images
If you are interested in obtaining the image database (it's basically a scaled-down version of a subset of LabelMe), send me an email and I'll put a .zip file online for you to download).