Abstract
Motion estimation algorithms are typically based upon the assumption of brightness constancy or related assumptions such as gradient constancy. This manuscript evaluates several common cost functions from the motion estimation literature, which embody these assumptions. We demonstrate that such assumptions break for real world data, and the functions are therefore unsuitable. We propose a simple solution, which significantly increases the discriminatory ability of the metric, by learning a nonlinear relationship using techniques from machine learning. Furthermore, we demonstrate how context and a nonlinear combination of metrics, can provide additional gains, and demonstrating a 44% improvement in the performance of a state of the art scene flow estimation technique. In addition, smaller gains of 20% are demonstrated in optical flow estimation tasks.