Refining 3D models using a two-stage neural network-based iterative process

This paper presents a refinement method that supplements the 3D model construction process. The refinement method addresses the issue of using inaccurate 3D positional information to construct the 3D model. In the context of this paper, the inaccuracies in the 3D information come from a low-cost and low-precision range finder system. The core component of the refinement system is a neural network architecture termed IFOSART that attempts to associate particular corrections to the 3D model given range and intensity information. Results presented show the refinement system successfully reduces the inaccuracies in real-world 3D models.