Deep-learning for 3D reconstruction

Depth perception is paramount for many computer vision applications such as autonomous driving and augmented reality. Despite active sensors (e.g., LiDAR, Time-of-Flight, struc- tured light) are quite diffused, they have severe shortcomings that could be potentially addressed by image-based sensors. Concerning this latter category, deep learning has enabled ground-breaking results in tackling well-known issues affecting the accuracy of systems inferring depth from a single or multiple images in specific circumstances (e.g., low textured regions, depth discontinuities, etc.), but also introduced additional concerns about the domain shift occurring between training and target environments and the need of proper ground truth depth labels to be used as the training signals in network learning. Moreover, despite the copious literature concerning confidence estimation for depth from a stereo setup, inferring depth uncertainty when dealing with deep networks is still a major challenge and almost unexplored research area, especially when dealing with a monocular setup. Finally, computational complexity is another crucial aspect to be considered when targeting most practical applications and hence is desirable not only to infer reliable depth data but do so in real-time and with low power requirements even on standard embedded devices or smartphones. Therefore, focusing on stereo and monocular setups, this thesis tackles major issues affecting methodologies to infer depth from images and aims at developing accurate and efficient frameworks for accurate 3D reconstruction on challenging environments.