This package contains the bag-of-colors signature extracted published in

| Christian Wengert, Matthijs Douze and Herve Jegou, | Bag of colors for improved image search | Proc. ACM Multimedia'11, November, 2011.

It also provides an evaluation framework to evaluate the global signature extraction.

There are several paramaters (number of colors, normalization type's, distance type, etc) that can be changed. The default is for a codebook comprising 256 colors and with almost optimal parameters with respect to the search accuracy.

CONTENT

The purpose of the files in the package root directory is as follows:

config_boc.m The configuration file, that contains both the library pathes and parameters for the descriptor generation.

compute_main_colors.m Extract the main colors using the procedure detailled in the paper. Note that this step is costly. Therefore the output of this stage (performed on the Flickr60K dataset) is already provided and stored in the matlab file 'maincolors.mat'

learn_color.m Learn the color codebook using the maincolors

compute_color_histo.m Function that compute the histogram of colors from an input image

boc_from_histo.m Convert an histogram of color into the descriptor

launch_compute_boc.m compute the set of BOC for a given dataset

evaluate_boc.m Compute the retrieval score

make_ppmsmall.sh This script was used to reduce the size of the images. Since we provide the images from Holidays already resized, you should not need it. It is required only if you want to consider a new dataset.

The subdirectories contain the following content:

utils/ contains some basic functions, evaluation functions, and configuration files for the dataset. Changing the pathes in the config_* files if needed.

flickr60k/ In this directory, we provide some data extracted from our Flickr60K dataset. It is used for the learning stage only. We provide pre-computed dominant colors, color palettes. The images are not provided itself, due to copyright reasons, but any large enough dataset of images from Flickr could do the job.

holidays/ This directory contains the ground-truth used to compute the score. It will also contain the re-sized images downloaded in the additional package.

PRE-REQUISITES

The yael library, more precisely with a working Matlab interface, is required to have a working and efficient execution of this package

Install a recent yael library. The last version can be obtained (with open-source license) at

https://gforge.inria.fr/projects/yael/

If you want to do all steps yourself, the INRIA holidays dataset is required.
You can download the set at

http://pascal.inrialpes.fr/data/holidays/jpg1.tar.gz and http://pascal.inrialpes.fr/data/holidays/jpg2.tar.gz

Use the make_ppmsmall.sh script to resize and preprocess the images.

INSTALLATION AND EXECUTION

First, you have to download and install Yael. Once you get a working yael, you should adapt the config_boc.m M-file to adapt the pathes to your configuration, and to the desired parameters.

The package must be used with some data that are provided in a separate archive. Download it from our website:

wget https://github.com/downloads/kooaba/bag-of-color/bag_of_colors_package.zip

and extract it in the root directory of the package. This additional package contains the re-sized and normalized images (created with make_ppmsmall.sh).

Then, you should be able to execute the following matlab commands to produce and evaluate global BOC descriptors.

learn_color % Learn a color codebook % This stage is not mandatory, as we already give % some pre-computed codebooks

launch_compute_boc % Compute the histogram of colors and BOC vectors

evaluate_boc % Compute the mAP value (by default, on Holidays)