Supplementary Material: Deep Discrete Flow

In this supplementary document, we first present additional experiments to show the convergence of training in Section 1.1 and compare the performance of exact matching against a frequently used approximate nearest neighbour search library in Section 1.2. We then show the speed-ups we gain by the strided implementation of the patch-based dilated convolutions in Section 2.1. We compare the runtime of forward propagating a full resolution image pair for the different local and context architectures in Section 2.2 and then report the runtime for individual stages of our pipeline compared to Discrete Flow in Section 2.3. For reproducibility, we list the parameter settings of the different stages in Section 3. Finally, we show additional qualitative results on the MPI Sintel and KITTI validation sets in Section 4. 1 Additional Experiments