We will present our line of work in the space of recognizing and tracking the
6D pose of object instances. We start off with a system that recognizes the 6D
pose of 3D objects from a single RGBD image. Current state-of-the-art systems
are
template based and capture the global appearance of objects. In contrast we
propose a dense patch-based combination of a discriminative method with a
generative model. The generative model leads to exact pose estimates while the
discriminative method vastly reduces the search space by efficiently producing
good proposals. We further show that results of the generative model can be
improved by incorporating the discriminative output. In the second part of the
talk we use that model for 6D pose tracking in a particle filter framework.
Here we combine the discriminative method's ability to produce proposals with a
motion model to allow real-time processing.