Image matting is one of the most strenuous but vital processes in filmmaking. It consists of extracting elements from different scenes, so they can be combined into a new, composite shot. With the finished shot, filmmakers can place actors in environments and situations that can’t be created physically on set. However, the matting process is a very mechanic and resource-intensive one, since it requires the skills of professional artists to achieve a clean, polished result. Factors like the number of elements, detail, and shot length can make this process more difficult. Matting consistently requires the use of multiple digital tools and labor hours, becoming a bottleneck in the post-production pipeline.

Even with the considerable amount of research in developing tools to ease this process, there is still a requirement of significant manual input from the artists. It is necessary to devise a way that produces a matte suitable for VFX applications with minimal user input. A compelling option is to look at the computer vision field. This industry is well known for having developed algorithms that can automatically detect and segment objects within an image, through the use of deep learning techniques.

This paper will present a deep learning-based algorithm that can perform automated rotoscoping of people on a given scene without any user input. 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, movement, and subjects. Automated rotoscoping is apt to be integrated into a VFX pipeline and will help artists be more efficient in their work.