High-fidelity, directable animation transfer using facial decomposition on optimized micro-sequences

娱乐产业向世界各地讲大量不同语言的人们提供媒体产品。媒体产品，如电视或网络广告、电影和视频游戏，通常由视频和声音流组成。为了适应不同的语言，制作人通常会在后期制作阶段用不同的语言重新生成声音流，这个过程被称为配音。虽然重新生成声音流（配音）仍然需要付出很多努力，但从研究和生产的角度来看，仍然存在的挑战之一是如何将原始视频流与新的音频流同步。由于演员的嘴唇动作与配乐不一致，这种不同步的效果会让人觉得很奇怪。在这个项目中，我们提出了一个实际的解决方案来解决配音同步问题，产生可以在制作工作室中使用的自然结果。为了实现这一目标，我们与位于蒙特利尔的制作工作室AudioZ建立了战略合作伙伴关系，我们将与他们在这个项目上进行合作。我们提出了一种视频合成方法，在保留原演员面部表情的同时，将动画集中在嘴唇区域。该方法采用新颖的优化技术，避免了视频合成中常见的恐怖谷效应。

The entertainment industry delivers media products around the world to people speaking a large number of distinct languages. Media products such as TV or internet spots, movies and video games, typically consist of a video and sound stream. To accommodate different languages, the producers typically re-generate the sound stream in different language in a post-production step, a process known as dubbing. While it still takes a lot of effort to re-generate the sound stream (dubbing), one of the remaining open challenges both from a research and production point of view is how to synchronize the original video stream with the new audio stream. The lack of synchronization creates a very uncanny effect that takes away from the experience since the lip motion of the actors does not correspond to the soundtrack. In this project we propose a practical solution to the dubbing synchronization problem, producing natural results that can be used in a production studio. To achieve that, we created a strategic partnership with AudioZ, a production studio located in Montreal with whom we will collaborate on this project. We propose a video synthesis method that blends animation focusing on the lip area while preserving the facial expression of the original actor. Our method uses novel optimization techniques to avoid the common undesired uncanny valley effect of video synthesis.