Multi-view 3d reconstruction using NeRF-based approaches

The growing demand for realistic and immersive 3d reconstructions has fueled advancements in computer vision and machine learning. This master thesis explores the usage of Neural Radiance Fields (NeRF) to achieve state-of-the-art 3d reconstruction from plain images. This research consists in an in-depth analysis of existing 3d reconstruction methodologies based on NeRF, highlighting their strengths and limitations. In particular, both the geometry and the texture are taken into account in analyzing the quality of the resulting meshes. After the analysis of the first results, NeuS and similar NeRF variants are explored in order to improve the smoothness and consistency of the mesh geometry. Finally, modifications of the texturing algorithm and the development of optimization procedures based on differentiable rendering are explored in order to obtain a more detailed and cleaner mesh texture. In conclusion, this master thesis contributes to the evolving field of 3d reconstruction by presenting a survey of the main NeRF-based methodologies and by showcasing the synergistic capabilities of NeRF, NeuS and other techniques such as differentiable rendering. The proposed methodology advances the state-of-the-art allowing for a more consistent mesh surface and a higher quality mesh texture.