Temporally Consistent Virtual Camera Generation from Stereo Image Sequences

摘要

The recent emergence of auto-stereoscopic 3D viewing technologies has increased demand for the creation of 3D video content. A range of glasses-free multi-viewer screens have been developed that require as many as 9 views generated for each frame of video. This presents difficulties in both view generation and transmission bandwidth. This paper examines the use of stereo video capture as a means to generate multiple scene views via disparity analysis. A machine learning approach is applied to learn relationships between disparity generated depth information and source footage, and to generate depth information in a temporally smooth manner for both left and right eye image sequences. A view morphing approach to multiple view rendering is described which provides an excellent 3D effect on a range of glasses-free displays, while providing robustness to inaccurate stereo disparity calculations.