I. Introduction
Pose estimation, or estimation of the 6 degree-of-freedom geometrical state from a single 2D image is an important problem that has received considerable attention over the years [1], [2], [3], [4], [5], [6], [7]. Applications include industrial automation such as bin picking (see figure 1), support systems for augmented reality as well as a whole range of consumer products including toys and house-hold appliances. Important properties of a real-world system for pose estimation is robustness to occlusion, changes in scale, and lighting. Occlusion is usually handled by using local descriptors [8], [9], and robustness to scale is usually solved by some kind of scale-space approach [10]. Robustness to lighting changes seems to be the most challenging problem, as will be made evident in the experiments section, and most local descriptors attempt to deal with this by using normalised features based on derivatives of the intensity.