The study of motion estimation reaches back decades and has become one of the central topics of research in computer vision. Even so, there are situations where current approaches fail, such as when there are extreme lighting variations, significant occlusions, or very large motions. In this thesis, we propose several approaches to address these issues. First, we propose a novel continuous optimization framework for estimating optical flow based on a decomposition of the image domain into triangular facets. We show how this allows for occlusions to be easily and naturally handled within our optimization framework without any post-processing. We also show that a triangular decomposition enables us to use a direct Cholesky decomposition to solve the resulting linear systems by reducing its memory requirements. Second, we introduce a simple method for incorporating additional temporal information into optical flow using "inertial estimates" of the flow, which leads to a significant reduction in error. We evaluate our methods on several datasets and achieve state-of-the-art results on MPI-Sintel. Finally, we introduce a discrete optimization framework for optical flow computation. Discrete approaches have generally been avoided in optical flow because of the relatively large label space that makes them computationally expensive. In our approach, we use recent advances in image segmentation to build a tree-structured graphical model that conforms to the image content. We show how the optimal solution to these discrete optical flow problems can be computed efficiently by making use of optimization methods from the object recognition literature, even for large images with hundreds of thousands of labels.