• Material Transfer
• Photogrammetry

Three-dimensional models are used in a large variety of different contexts. However, manually creating 3D assets from scratch is a time-consuming process. Acquisition techniques have become popular for rapidly obtaining 3D assets, and among these, photogrammetry is a technique that allows one to reconstruct 3D models from multiple acquired photographs of an object. The material of the acquired object can have a non-negligible reflective component which can lead the photographs to contain strongly lit areas such as highlights or reflections.
In this thesis, we propose a method for improving photogrammetry by processing the acquired photographs. To achieve this goal, we generated a synthetic dataset of rendered images that simulated different photogrammetry setups for the acquisition of various models and materials. We used this dataset to experimentally validate the assumption that the reconstruction quality of a model depends on the material being photographed. We observed that reflective materials compromise the quality of photogrammetric reconstructions. Therefore, we developed a data-driven approach to apply material transfer on input images. Using this approach, it was possible to obtain dull materials from acquired reflective materials, thereby reducing the impact of specular reflections on the 3D reconstructed model. This resulted in more detailed reconstructions.