Consistent 2D-to-3D video conversion using spatial-temporal nonlocal random walks

We propose a novel spatial-temporal nonlocal random walks (SPNRW) algorithm to generate consistent and smooth depth maps for input video sequences. Users just add some scribbles on key frames at regular interval, which is based on the estimated depth of the whole video sequence. Since the proposed spatial-temporal nonlocal random walks can preserve the structure of boundaries between frames effectively, it will also generate consistent dense depth maps of the video after introducing the spatial-temporal information. Experimental results on video sequences demonstrate that our method can obtain better consistent and accurate dense depth maps.