Research On 2D To 3D Conversion Methods Based On GIST Feature Matching And SIFT Flow

With the increase of 3D silver screens and the popularity of 3D hardware products, existing 3D media resources are far from meeting the growing market demand. In order to solve the problem and promote the development of 3D industry, people try to convert 2D media resources into stereoscopic ones, such as movies and TV shows taken by professional directors, and common photos and videos taken in daily life. Those 3D results can be displayed in the projection terminal so that the audience can watch them with 3D glasses.This thesis proposes an automatic 2D to 3D conversion method. Based on the analysis of the scene similarity between different images, several candidate images are selected from the RGBD database by implementing feature matching and a mapping relationship is built between the input image and one of the candidate images. Then utilizing the known depth information, a probability model is established to estimate the depth image of the input image. Finally, it can be realized to synthesize 3D image. This thesis follows three aspects in detail:Firstly, taking advantage of GIST feature matching method, candidate images from the database whose scenes are similar with the input image are selected. To achieve a more comprehensive understanding of the known information, GIST features of each image are calculated and the result can be transferred into a high dimensional feature vector. According to the distance between different vectors, the correspondingly candidate images can be picked out from the database.Secondly, this thesis manage to build many-to-one mapping relations of pixels between the input and candidate images depending on SIFT flow method in order to complete the assignment and optimization of the depth. Computing the SIFT features for every pixel, the SIFT image can be obtained as a descriptor consisting of all the features. Then, SIFT flow vector is able to be acquired by solving energy function between two SIFT images, which records the desired mapping relationship.Thirdly, a probability model is established to describe depth distribution of the input image, which contains information of the input, the mapping relation corresponding to the matching images, gradient information and mean information. Then Iterative Reweighted Least Squares method is used to estimate depth and attain the precise depth map. Finally, use the original 2D image and the precise depth map to generate the corresponding left view and right view, and the ultimate 3D image.