3D display has become the inevitable trend of display technology. Converting the traditional and classical 2D videos to 3D videos is an important and effective measure to solve the shortage of 3D contents. The major work about 2D-to-3D video conversion is how to extract depth information from the 2D video, and synthesize a new image from the existing viewpoint. We propose a depth extraction method based on dense edge-preserving optical flow from 2D videos, reducing the matching error in textureless regions. Moreover, we use the Gaussian Pyramid and Laplace Pyramid at cross scales to fill the holes in the image at new view point after 3D warping. The experiments show that our results outperform state-of-the-art methods in visual effect and statistics.