Automatic Depth Estimation for Indoor and Outdoor Scenes

• Main Authors: Kuei-Da Peng, 彭奎達
• Other Authors: Cheng-Chin Chiang
• Format: Others
• Published: 2013
• Online Access: http://ndltd.ncl.edu.tw/handle/27144084400537002534

碩士 === 國立東華大學 === 資訊工程學系 === 101 === In recent years, 3D contents has become more and more popular and leads to the new trend in movie and game industry. Many 3D viewing peripherals including the red/blue glasses in earlier days and the polarized shutter glasses nowadays, have become more and more common to promote the new market of 3D visuals. However, a more important key technology that can prevail the market would be the technique of converting 2D visual contents to 3D scenes. This thesis aims to propose a novel solution that can do such conversions automatically. In this thesis, we develop an automatic depth estimation method for both outdoor and indoor scenes. Given an input 2D image, the proposed method performs a content-based image retrieval over a database of images. The retrieval then returns 30 images which are similar to the input image in color histograms. Associated with each retrieved image, its corresponding depth image is also returned. By applying the SIFT flow algorithm to reconfigure both the layout of each image and its corresponding depth image, we obtain a new set of candidate images which reveal better visual consistency with the input image. By partitioning the input image into four overlapped rectangular regions, the proposed method further performs region-based matching to choose the seven most similar regions among the reconfigured images for each of its four regions. Consequently, the depth information of the chosen most similar regions can be used to estimate the depth of the region on the input image. The depth is roughly estimated by applying the median filter. Afterwards, we apply the cross bilateral filter to smooth the depths of pixels, while preserving the edge property well over the estimated depth image. Finally, we synthesize a red-blue 3D image for the 2D input image such that we can view the 3D stereo visuals via a red-blue glasses. The experiments show that the proposed depth estimation method outperforms other prior methods