3d Stereoscopic Image Generation from Monocular View in Structured Environments

With the rapid growing need for 3D image generation, distribution and display as well as the large repository of 2D images in the Internet, creating 3D stereoscopic images from 2D images has becoming an important and urgent issue in recent years. The key step in conversion from 2D to 3D is depth map generation. For the step of depth map generation, most existing methods are based on focus & blur cues or vanishing point detection. In this work, we propose a novel depth map generation method based on plane extraction and demonstrate that the plane extraction step is crucial for accurate and reliable depth map generation. The main idea of the proposed plane extraction algorithm is to detect corners and edges simultaneously based on autocorrelation function of the 2D image with optimized parameters with the aid of kernel density theory. After extracting the foreground and partitioning the background into several planes, it is very convenient to accurately estimate a relative depth map with the cues of planar geometric features. The depth map is then used to generate the stereo pair before obtaining the final image with 3D effect. Experimental results validate the superior performances of the proposed method with comparison of the state-of-art method.