Recovering Local Geometry for Time-varying Mesh from Motion

3D geometric modeling of real-world characters from multiview video is one of the popular topics in computer vision. In order to capture detailed and accurate 3D geometric model (Time-varying mesh), however, most proposed systems use only some static components(e.g. intensity, texture, color) of the generalized frames, e.g. photometric stereo method, stereo matching method. On the other hand, the dynamic components (e.g. optical flow field), are usually used for motion tracking but not for surface geometry capture. Actually, it is also possible to derive surface orientation and to characterize certain types of surface discontinuities by their motion [1]. That is, the processing of motion tracking and surface geometry capture can be done at the same time if we used the dynamic components information of the frames, which can be quite useful in 3D video compression. Therefore, it's necessary to develop a method to apply recovering local geometry from motion to TVM modeling. In this paper, we propose a new method to recover local geometry for TVM from motion based on recent advances in dense optical flow algorithms.