Semi-Automatic Depth Map Generation in Unconstrained Images and Video Sequences for 2D to Stereoscopic 3D Conversion

Abstract

In this work, we describe a system for accurately estimating depth through synthetic depth maps in unconstrained conventional monocular images and video sequences, to semi-automatically convert these into their stereoscopic 3D counterparts. With current accepted industry efforts, this conversion process is performed automatically in a black box fashion, or manually converted using human operators to extract features and objects on a frame by frame basis, known as rotoscopers. Automatic conversion is the least labour intensive, but allows little to no user intervention, and error correction can be difficult. Manual is the most accurate, providing the most control, but very time consuming, and is prohibitive for use to all but the largest production studios. Noting the merits and disadvantages between these two methods, a semi-automatic method blends the two together, allowing for faster and accurate conversion, while decreasing time for releasing 3D content for user digest. Semi-automatic methods require the user to place user-defined strokes over the image, or over several keyframes in the case of video, corresponding to a rough estimate of the depths in the scene at these strokes. After, the rest of the depths are determined, creating depth maps to generate stereoscopic 3D content, and Depth Image Based Rendering is employed to generate the artificial views. Here, depth map estimation can be considered as a multi-label image segmentation problem: each class is a depth value. Additionally, for video, we allow the option of labeling only the first frame, and the strokes are propagated using one of two techniques: A modified computer vision object tracking algorithm, and edge-aware temporally consistent optical flow./p pFundamentally, this work combines the merits of two well-respected segmentation algorithms: Graph Cuts and Random Walks. The diffusion of depths, with smooth gradients from Random Walks, combined with the edge preserving properties from Graph Cuts can create the best possible result. To demonstrate that the proposed framework generates good quality stereoscopic content with minimal effort, we create results and compare to the current best known semi-automatic conversion framework. We also show that our results are more suitable for human perception in comparison to this framework.