Realistic video avatar

In this paper, we present a system for the implementation of a photorealistic avatar using video captured from a user. This is achieved by constructing the dynamic video texture map and combining it with the 3D mesh model of the user to render the photorealistic avatar. The dynamic video texture map reflects the user’s facial expressions and is generated by composing the cylindrical texture map of the user with constant updates either directly from the input video or derived from a temporal and spatial interpolation scheme. To derive the 3D mesh model for the user, we define feature points on a generic model and use the feature points identified on the video sequence to deform and obtain the avatar model. The goal of this realistic video avatar project is to provide a vivid representation of participants with a more realistic quality avatar, as compared to using only the facial animation parameters defined in MPEG-4 without the corresponding image updates. This scheme is suitable for conferencing applications because it requires much lower bandwidth than live video, and yet provides a 3D avatar representation for any virtual environment. 1