Recent advances in 3D have increased the importance of stereoscopic content creation and processing. Therefore, converting existing 2D videos into 3D videos is very important for growing 3D market. The most difficult task in 2D-to-3D video conversion is estimating depth map from single-view frame images. Thus, in this paper, we propose a novel motion-based 2D to 3D video conversion method. The method first determines the motion type using the optical flow estimation. Then, different depth estimation processes are performed based on the motion type. For global motion, the depth from motion parallax provides the final depth map. For local motion, the depth from template together with the bilateral filter is used to produce the depth map. Finally, the left- and right-view images are synthesized to generate realistic stereoscopic results for viewers. During the process, the visual artifacts of the synthesized virtual views are effectively eliminated by recovering the separation and loss of foreground objects. A comparative study and quantitative evaluation with other conversion methods are carried out, which demonstrate that better overall quality results may be obtained using the proposed method.
Pattern-based Depth Map Generation for Low-complexity 2D-to-3D Video Conversion