Lecture

Having perceptual color differences is key in many computer vision applications
such as image segmentation or visual salient region detection. As the RGB
spaces are known to be perceptually non uniform, some transformations have been
introduced in the literature to map RGB coordinates into uniform spaces such
as CIELAB or CIELUV. However, it is worth noting that these transformations
require the knowledge of the acquisition conditions such as the illumination
or the acquisition device. When these conditions are unknown, as in most of
the applications, the distances evaluated in the resulting spaces are only
roughly perceptual.
In this task, I will present a new local Mahalanobis-like metric learning
algorithm that aims at accurately approximating a perceptual color difference
that is invariant to the acquisition conditions. Using the theoretical
framework of uniform stability, I will present consistency guarantees on the
learned metrics and show that they outperform the state of the art in two
experimental settings: the first one aims at assessing the ability of the
learned metrics to generalize to new colors and devices, while the second
makes use of them in a segmentation application.