Video matting is a problem of separating foreground object from background in a video sequence. We developed an algorithm that takes two fully segmented keyframes as an input and generates alpha mattes (opacity channels) for all intermediate frames by means of optical flow.
Our main goal is visual quality of the result when the extracted object is composited onto new background. The algorithm saves user’s time as compared to manual processing of frames (rotoscoping). The result is fully controllable in a non-parametrical way by keyframes and additional strokes, which allows achieving desirable accuracy.
This project is supported by Microsoft Research programs in Russia.