Abstract
Computer vision based localization and pose estimation of known objects within camera images is often approached by optimizing some sort of fitting cost with respect to a small number of parameters including both pose parameters as well as additional parameters which describe a limited set of variations of the object shape learned through training. Gradient descent based searches are typically employed but the problem of how to "weigh" the gradient components arises and can often impact successful localization. This paper describes an automated, shape-adaptive way to choose the parameter weighting dynamically during the fitting process applicable to both standard gradient descent or momentum based accelerated gradient descent approaches.