Improved Integration of Depth Images in 3D Models

Dennis Depner

Supervisor(s): Stefan Ohrhallinger

Vienna University

Abstract: 3D reconstruction using depth images is established in many areas of computer graphics, such as virtual and augmented reality. Our research is about the procedure for 3D reconstruction from the Kinect Fusion algorithm. Particularly, the work aims to determine how the inclusion of the aggregated point cloud information of previous / older depth images as median vectors affects the quality of the 3D reconstructed surface. In addition to this, two algorithms for the generation of meshes are compared, which are the Marching Cubes algorithm and an adaptation of it, which can be specifically applied to octrees. The results have shown that including median vectors has a positive influence on the quality of the 3D reconstructed surface compared to the default Kinect Fusion algorithm. In smaller areas, there are fewer holes on the surface. Finer objects also appear softer and are reconstructed less edgy. The adapted Marching Cubes algorithm for octrees also shows a strong improvement in terms of generating fewer holes in the mesh.
Keywords: 3D Reconstruction, Computational Geometry, Geometry Processing
Full text:
Year: 2022