A practical approach to view-based synergetic pose estimation

This paper reports on progress towards a system that will track the pose (orientation) of a rapidly moving rigid object. The need for high-speed pose estimation led us to use a view-based, rather than a CAD model-based approach. In our work, we build a view-based model of the object using a training set of images with known pose. We compress the training set data using synergetic feature extraction, which allows us to store our model in the form of a relatively low-dimensional hyper-surface parameterised by the pose variables. Our model is stored compactly and can be evaluated quickly to provide estimates of pose for novel images. Our results reveal low mean-error values that are resistant to high levels of unstructured noise and low levels of structured noise