Watershed and Random Walks based depth estimation for semi-automatic 2D to 3D image conversion

Depth map estimation from a single image is the key problem for the 2D to 3D image conversion. Many 2D to 3D converting processes, either automatic or semi-automatic, are proposed before. Quality of the depth map from automatic methods is low and there exists wrong depth values due to errors estimation in depth cue extraction. The semi-automatic approaches can generate a better quality of depth map based on the user-defined labels, which indicate a rough estimation of depth values in the scene, to generate the rest of depth value and reconstruct the stereoscopic image. However, they require complexity system and are very computational intensive. A simplified approach is to combine the depth maps from Graph Cuts and Random Walks to persevering the sharp boundary and fine detail inside the objects. The drawback is the time consuming of the energy minimization in the Graph Cuts. In this paper, a fast Watershed segmentation based on the priority queue, which indicates the neighbor distance relationship, is used to replace the Graph Cuts to generate the hard constraints depth map. It is appended to the neighbor cost in the Random Walks to generate the final depth map with hard constraints in the objects boundaries regions and fine detail inside objects. The Watershed and Random Walks are low computational intensive and can achieve approximate real time estimation which results in a fast stereoscopic conversion process. Experimental results demonstrate that it can produce good quality stereoscopic image in very short time.