New editing techniques for video post-processing

摘要

This thesis contributes to capturing 3D cloth shape, editing cloth texture and altering object shape and motion in multi-camera and monocular video recordings. We propose a technique to capture cloth shape from a 3D scene flow by determining optical flow in several camera views. Together with a silhouette matching constraint we can track and reconstruct cloth surfaces in long video sequences. In the area of garment motion capture, we present a system to reconstruct time-coherent triangle meshes from multi-view video recordings. Texture mapping of the acquired triangle meshes is used to replace the recorded texture with new cloth patterns. We extend this work to the more challenging single camera view case. Extracting texture deformation and shading effects simultaneously enables us to achieve texture replacement effects for garments in monocular video recordings. Finally, we propose a system for the keyframe editing of video objects. A color-based segmentation algorithm together with automatic video inpainting for filling in missing background texture allows us to edit the shape and motion of 2D video objects. We present examples for altering object trajectories, applying non-rigid deformation and simulating camera motion.