오더 모션기반 대화형 안면 애니메이션 합성

초록

Talking face animation consists of generating a video so that the face in a source image is animated according to the order motion of a driving video. Our framework solves this problem without using prior information and annotation about specific subjects to generate animation. Once trained on a set of driving videos showing objects of the same category, the model could be applied to any object of this class. We separate motion and appearance information using a self-supervised formulation to achieve this. We use representations containing a set of learned keypoints with their local affine transformations to maintain complex motions. A generator net models occlusions coming during target motions and merge the appearance taken out of a source image and a motion from a driving video.