A Tree Parameterization for Efficiently Computing Maximum Likelihood Maps using Gradient Descent