A Research on 3-Dimensional Image Generation from Continuous Movie Frames

使用视差视图的3D立体图是一种展示图像真实感和氛围的技术。虽然企业开发了很多以这种方式为基础的显示设备，但必须为该设备开发新的软件内容。使用视差视图的立体图显示不依赖于场景的完整3D信息。它只需要两张视差图像。由于该方法的性质，该方法比传统的三维模型重建技术需要更少的计算量。本研究的目的是建立由单镜头摄像机拍摄的连续视频帧产生视差图像的技术。我们用两种不同的方法进行了这项研究。一种方法是基于连续视频帧计算的光流数据进行深度估计。如果视频帧是通过摄像机的平行运动来拍摄的，那么像素的位移量（光流）就对应于像素的深度。利用光流估计的深度信息，可以计算出两幅视差图像。我们开发了该方法的基本算法和实验系统。另一种方法是基于改变视点的视频帧的体空间重建。该方法对场景的三维结构和表面纹理进行重构。该方法比基于光流的方法更具潜力。然而，静态空间约束适用于客观场景。

3D stereogram using parallax views is a technique to demonstrate realism and ambiance of images. Many display devices based on this method are developed by companies, but software contents must be created newly for the device. Stereogram display using parallax views does not depend on complete 3D information of the scene. It needs only two images with parallax. As the nature of this methods, this method requires less computation than conventional 3D model reconstruction techniques. The goal of this research is to establish techniques to produce parallax images from continuous video frames taken by single lens video camera. We performed the research by two different approaches. One approach is based on depth estimation from optical flow data computed from continuous video frames. If video frames were taken by parallel movement of camera, the amount of pixel shift (optical flow) corresponds to the depth of the pixel. By using depth information estimated by optical flow, two images with parallax can be computed. We developed basic algorithm and experimental system for the method. The other approach is based on volume space reconstruction from video frames that is taken with changing viewpoint. The method reconstructs 3D structure and surface texture of the scene. This method is more potentiality than optical flow based approach. However static space constraint for the objective scene applies.