Natural-Textured Mesh Stream Modeling from Depth Image-Based Representation

摘要

This paper presents modeling techniques to generate natural-textured 3D mesh stream from depth image-based representation (DIBR). Although DIBR is a useful representation for expressing 2.5D information of dynamic real objects, its usage is limited to point-based applications. In order to generate smooth and textured 3D mesh models, depth images are captured using active depth sensors, and they are processed with segmentation, noise filtering, and adaptive sampling technique based on the depth variation. 3D meshes are reconstructed by constrained Delaunay triangulation and smoothened with the 3D Gaussian filter. Each mesh model is parameterized for texture mapping of a corresponding color image. Proposed procedures are automated to generate 3D mesh stream from hundreds of image sequence without user interventions. Final output is a natural-textured mesh model per frame, which can be used for arbitrary view synthesis in virtual reality or broadcasting applications.