Research Of RGB - D Geometric Feature Extraction And 3D Shape Recovery

With the advent of Kinect devices, research on RGB-D image is becoming one of the hot fields of computer vision. RGB-D image uses the 2D image to represent the 3D scene information. It plays an important role as a bridge between 2D and 3D. Meanwhile it is becoming possible to use the image processing method to solve 3D problems. The image feature has been one of the key issues for image processing. In this thesis, geometric features extraction and classification approach for RGB-D image and its application in intrinsic scene properties recovery was proposed. This algorithm can not only output reflectance, illumination and shading, but also optimize the raw depth image. The main contents and contributions of this thesis are as follows:First, geometric feature extraction classification algorithm for RGB-D image is researched in this thesis. By combining two dimensional image and three dimensional objects in depth image information, the object of geometry feature extraction and feature information classification is achieved. Using this kind of algorithm, the geometric features of the edge can be classified as occlusion, convex and concave. The classification result of geometric features is an advanced semantic feature that can be applied to other areas of computer vision. NYU Depth Dataset is used for testing and analysis. Experimental results showed that the algorithm can extract and classify the geometric features clearly, completely and accurately with high robustness.Second, an edge-based Scene-SIRFS intrinsic scene properties recovery algorithm was proposed, which is an application for the classification results of the geometric features. The edge-based Scene-SIRFS intrinsic scene properties recovery algorithm is an improved algorithm which is based on Scene-SIRFS algorithm. The algorithm related to the nature of the occlusion edges and the fold edges, as well as the introductions of the occlusion edges and the fold edges priori constraints, using minimal optimization approach to achieve more accurate intrinsic scene properties recovery. In this thesis, the NYU Depth Dataset is used again for testing and analysis. The experimental results are compared with the results of Scene-SIRFS algorithm during the experiment. The experimental results showed that the result of edge-based Scene-SIRFS intrinsic scene properties recovery algorithm is better than that of Scene-SIRFS algorithm. The parts of edges are clearer, and the geometric features are more obvious. The refining results of the original depth map are more accurate.