Rotoscope Automation with Deep Learning

摘要

We present a deep learning based algorithm that can automatically rotoscope people on a given scene, without any user input. Current approaches to image matting require a significant amount of human input, whether by manually rotoscoping or by doing a chroma key. We show that this algorithm can perform comparably and even surpass the rotoscoping capabilities of After Effects' RotoBrush tool, in a variety of scenes comprising different lighting conditions, movements, and subjects. This makes it suitable for an integration within a VFX pipeline.