Specular Normal Synthesis Using Stochastic Super-resolution for Detailed Facial Geometry
Abstract
Detailed facial geometry is critical for the visual realism of face models in computer games, movies, and virtual reality applications. The existing face scanning methods, however, are either sacrificing resolution for real-time processing, or requiring expensive high-speed cameras. In this work we propose a new technique for real-time high-resolution facial scanning using spherical gradient illumination. The key elements of the approach are the use of stochastic super-resolution to generate specular normal map based on diffuse normal map, instead of capturing both of them during scanning process. We analyze a training dataset of diffuse normal maps and specular normals of a particular object and learn the mapping from low-frequency components of diffuse normal maps to high-frequency components of specular normal maps of that object. This enables us to infer, for example, the most likely high resolution specular normal map detail depicting the same person as a low-resolution diffuse normal map given as input. Experimental results show that the proposed algorithm generates high-quality specular normal maps from diffuse normal map inputs.
Document Details
- Document Type
- Technical Report
- Publication Date
- Jan 01, 2010
- Accession Number
- AD1171502
Entities
People
- Abhijeet Ghosh
- Jun Zheng
Organizations
- University of Southern California