Beyond Depth-Camera Domain Shift: an analysis on 3D visual recognition across different sensors

3D cameras are crucial for many robotic tasks as object detection, pose estimation and scene understanding. Since the advent of Microsoft Kinect, many other depth sensors have been developed and artificial agents have been endowed with different systems depending on the specific requirements. Due to the variety of acquisition logics (structured light, time-of-flight, active stereo), the obtained depth data differ significantly which limits the possibility to export knowledge and learned models across different platforms. The aim of this thesis is to run an extensive analysis on current existing deep learning model to check if and how their performance is affected when changing the depth data domain. We will also study how to integrate them with domain adaptation strategies to alleviate the existing distribution shift and allow knowledge transfer.