We propose an architecture based on Convolutional Neural Networks (CNNs) for the detection of motion boundaries from two consecutive images of a video sequence. Existing learning-based approaches start with dense optical flow estimates, which are expensive to compute and often fail near motion boundaries, exactly where they are needed most. In contrast, we explore ways to detect motion boundaries without first computing optical flow. For efficiency, we hypothesize that motion boundaries occur at or near the edges of superpixels in an over-segmentation of the first image.

Providing a satisfactory quality of experience (QoE) to Internet users is crucial for content and service providers. When users get bad QoE from an application, such as the videos they are watching on a streaming provider keep freezing or the shopping Web site they are visiting takes a long time to load, they often spend less time on the application, return to it less frequently, or even worse they might switch to an alternative application, in all cases hurting the business financially.