Color Descriptors

Photometrically Robust Color Descriptors

mug shot Although color is commonly experienced as an indispensable quality in describing the world around us, many feature-based representations are only based on shape description, and ignore color information. The description of color is hampered by the large amount of variations which causes the measured color values to vary significantly. A change in illuminant color, viewpoint, and acquisition material, all influence the color values of the scene. To accomplish a wide applicability of color description, it should be robust to : 1. photometric changes commonly encountered in the real world, 2. varying image quality, from high quality images to snap-shot photo quality and compressed internet images. Based on these requirements we derived a set of color descriptors.

The figure shows how the combined shape and hue descriptor is computed. The top row shows the computation of the SIFT descriptor. The bottom line shows the computation of the hue descriptor. The similarity between the two computations is evident from the figure. The SIFT describes the direction of the edges in a local patch. Four direction histograms are computed, where the gradient of the edge determines the strength of the update. For hue descriptor, we compute the hue and saturation at each position; these can also be represented as a vector, where the hue is the angle and the saturation the length. We compute the hue histogram of the patch where the strength of the update is equal to the saturation of the measurement. This ensures that pixels with low saturation (black-grey-white), where the hue is undefined, have no influence on the final color descriptor.

A tar-file containing the matlab code for the color descriptors can be downloaded from : ColorDescriptors.tar.

Image Classification Data Set

mug shot To test the color descriptors we collected a data set of soccer teams. This data set contains images from 7 soccer teams taken from the web, containing 40 images per class, divided into 25 training and 15 testing images per class. Although, players of other teams were allowed to appear in the images, no players being a member of the other classes in the database were allowed.

A tar-file containing the images: soccer_data.tar


Joost van de Weijer, Cordelia Schmid Coloring Local Feature Extraction, Proc. ECCV06, Graz, Austria, 2006.