Occlusion Reasoning for Temporal Interpolation using Optical Flow

We present an optical-flow-based algorithm to smoothly interpolate between two images. We reason about the depth ordering of objects, and show how bidirectional flow can be used to reduce holes in the estimated flow at the interpolated time and perform occlusion reasoning. We develop a purely pixel-wise algorithm and then add spatial regularization. We evaluate our algorithm on the interpolation set of the Middlebury flow benchmark.