Mesh-based Multi-view Normal Integration with Energy Minimization Using Surface Reflectance Properties

Abstract

We propose a technique to reconstruct a general 3D object using surface reflectance information from multiple viewpoints. Our core optimization framework uses multi-view normal integration, which can recovers water-tight surface of the object iteratively in a coarse to fine manner. The integration requires normal vector field from multiple viewpoints, which we can derive from surface reflectance. We then handle the topological changes if self-intersection occurs from the optimization. We also employ the idea of multi-resolution and weighted data heuristic which helps dealing with noisy data and improves both accuracy and optimization time. Our experiment shows that the framework is able to robustly recover 3D surface well with both synthetic and real data.