GOING FROM ITOF TO DTOF: METHODS FOR MPI CORRECTION AND TRANSIENT RECONSTRUCTION

Depth data contains vital information for understanding the geometry of objects and placing them in the 3D world. To this aim, several sensors have been developed to retrieve depth images, each with its own strength and weaknesses. In this thesis we will cover Time-of-Flight cameras (ToF), a particular kind of sensor which computes the distance information based on the time of flight of a light impulse. We will mainly focus on indirect ToF cameras, which are commercially available devices working at interactive frame rates with until 1 Megapixel resolution. These devices are affected by Multi-Path Interference (MPI), a key challenge causing distortion in the depth estimation process. Common data-driven approaches tend to focus on a direct estimation of the output depth values, ignoring the underlying transient propagation of the light in the scene. In this Ph.D. thesis instead, we propose some very compact deep learning architectures for transient data estimation, the first one exploiting a two-peaks encoding of the transient, the second one leveraging on the direct-global subdivision of transient information. Afterwards, we will instead deal with ToF sensors in a power-constrained environment, such as that of mobile devices. In this setting, we will propose a quantized neural network for depth completion, which reaches competitive performance while keeping its size limited. Finally, we switch to hyperspectral images and see a work on learning architectures for spectrum reconstruction with limited amounts of ground truth data.