Single View Depth Estimation via RGBD Big Data

2D-to-3D conversion is one way to alleviate the lack of 3D-TV program material.The most important and difficult issue in 2D-to-3D conversion is how to estimate the depth map from a monocular image.This paper proposes a PatchMatch depth transfer method of depth estimation from a monocular image for 2D-to-3D conversion based on RGBD data from internet.First, the proposed method retrieves K-nearest neighbor images from RGBD database using global image descriptors such as GIST features.Then, the proposed method matches the input image to its neighbor images by the PatchMatch method.Third, the proposed method transfers depth maps of neighbor images to the input image and estimates its initial depth by median filtering on these transferred depth maps.Finally, the proposed method refines the initial depth map using tri-lateral filtering, in order to further improve the depth map estimation quality.Experimental results show that the proposed approach can greatly reduce the computation burden with depth map quality improvement compared to the scale invariant feature transform flow based method.