Learning detailed models from images and videos using machine learning techniques and applications to graphics

点击翻译按钮获取中文摘要

Synthetic images and animations of real scenes are now common place in different domains such as medicine, computer games, feature films, TV advertising, homeland security, and fine arts. Creating these images and animation is the first goal of computer graphics. Traditional computer graphics approaches start with geometric models and then generate and display virtual representations. Many computer graphics approaches have been successful. However, it is clear that synthetic images and animation still look artificial and that the cost and time have to be lowered. In the past, the field of computer graphics has been considered as the inverse of computer vision. Indeed, computer vision starts with input images and videos and process them to understand the geometric and physical properties of objects and scenes. The objectives of my research are the integration of computer vision and computer graphics techniques, and the creation of a framework in which these two domains collaborate through machine learning techniques to model the world around us (e.g. human body and motion, natural scenes, rigid and non-rigid objects) directly from measurements. These measurements will be learned from real images and scenes. Indeed, recent years have seen a significant technological development in the areas of high-quality sensors which have simplified the acquisition of the world content.