In this paper, we propose a novel method to generate highly photorealistic three-dimensional (3D) inner mouth animation that is well-fitted to an original ready-made speech animation using only frontal captured images and small-size databases. The algorithms are composed of quasi-3D model reconstruction and motion control of teeth and the tongue, and final compositing of photorealistic speech animation synthesis tailored to the original. In general, producing a satisfactory photorealistic appearance of the inner mouth that is synchronized with mouth movement is a very complicated and time-consuming task. This is because the tongue and mouth are too flexible and delicate to be modeled with the large number of meshes required. Therefore, in some cases, this process is omitted or replaced with a very simple generic model. Our proposed method, on the other hand, can automatically generate 3D inner mouth appearances by improving photorealism with only three inputs: an original tailor-made lip-sync animation, a single image of the speaker’s teeth, and a syllabic decomposition of the desired speech. The key idea of our proposed method is to combine 3D reconstruction and simulation with two-dimensional (2D) image processing using only the above three inputs, as well as a tongue database and mouth database. The satisfactory performance of our proposed method is illustrated by the significant improvement in picture quality of several tailor-made animations to a degree nearly equivalent to that of camera-captured videos.
ASJC Scopus subject areas
- コンピュータ サイエンス（全般）