Scene analysis based on an optimization of CG images

Most methods of image analysis are based on the bottom-up approach from the data to the model, because the problem is assumed as an inverse of CG rendering. Therefore, some hypotheses or limitations are imposed on the object, illumination or motion to cope with this ill-posed problem, and no practical but general framework is proposed. On the other hand, the techniques for rendering computer graphics are evolved in aspects of both quality and speed, and most optical phenomena can be simulated precisely in real-time. Under such background, we are engaged in a research of synthesisbased image analysis, such as an motion tracking system without feature extraction. Same framework can be applied to reconstruction of surface reflectance or illumination condition.