Algorithms For Depth Map Generation From Monocular Videos

With the great development of science and technology in recent years, more and more researchers begin to concern about the techniques of converting two dimension (2D) video to three dimension (3D) video. The paper begins with the introduction of the background and development of 3D video, and then provides a review of the latest development of 3D video. It also introduces stereoscopic display technique which makes use of monocular videos and depth information. Depth map generation is the main issue of that technique. The paper shows how to get the depth information only from monocular 2D video sequence, which provides feasible solutions for drawing the depth maps.The thesis proposes two semi-automatic methods in order to derive depth maps: Firstly, we propose an improved color adaptive depth extrapolation (CADE) filter to get depth information. Subsampling is used to improve the efficiency of the CADE filter while median post-filter and adaptive window are introduced to improve the quality of the CADE filter. Secondly, we present a semi-automatic depth map generation algorithm based on belief mapping, which acquires the depth information of key frame and bi-direct bilateral depth generation algorithm. When the de-occluded area has similar color with the nearby object area, belief mapping can reduce the significant blur for de-occluded areas. Finally, two methods are compared in performance.The conversion from 2D video to 3D video is a new and difficult research topic. It is closely related to computer vision, human visual system, image and video processing. The content of the thesis contributes to the generation of depth maps, and verifies their feasibility and efficiency through experiments.