An exploration of deep learning methods in mesh refinement for multi-view stereo

3D reconstruction is the process of capturing the shape and appearance of real objects. It is a complex and fascinating problem with a long history and a large set of effective algorithms. In this context, Multi-View Stereo Mesh Refinement is the process of refining a reconstructed mesh using a set of 2D calibrated images. The current state-of-the-art approach is based on the definition a suitable image-based similarity function. We look at Deep Learning methods, which have recently provided groundbreaking improvements in many sorts of image related tasks, to provide a more effective, data-based one which can overcome current unrealistic assumptions. Given the intrinsic geometric nature of the problem, we look at the recent breakthroughs in Geometric Deep Learning to search for a path to a more natural representation of the problem and a more efficient and accurate solution.

In questa tesi vengono illustrati due approcci al problema di raffinamento di una mesh dato un set di viste calibrate