3D Reconstruction of Indoor Scenes: a Neural Radiance Field Approach Supervised by Depth Priors

The target of this thesis is the geometric reconstruction of indoor scenes. We exploit an approach based on Neural Radiance Fields (NeRF) and able to learn an implicit representation of the scene geometry, starting from RGB images only. In particular, the geometry is learnt employing the Signed Distance Function (SDF) which estimates the closest surface for every point in the volume enclosing the scene. Starting from NeuS, an already existing method, we implement some supervision strategies to help the networks during the training phase. One approach performs a depth supervision based on the scene point cloud estimated by COLMAP, a Structure-from-Motion algorithm already used for camera pose estimation. The point cloud, obtained from features matched and triangulated over many views, adds sparse geometrical constraints to the geometry learnt by the model, thus increasing the reconstruction accuracy of complex structures. Then, we propose a novel depth supervision based on depth maps, in order to focus the NeRF learning on areas close to the surface. The last improvement is a color compensation strategy, to handle images acquired with variable exposure and white balancing settings. This leads to a more stable convergence and it helps the geometry estimation as well. Overall, the resulting method can produce accurate and colorful reconstructions of indoor environments. We test our method on indoor scenes, showing the effects of our implementations. In addition, we investigate the importance of testing on scenes acquired explicitly for NeRF based reconstruction, discussing the most important requirements to meet in the case of custom dataset acquisitions.