Prototype of SAILab Virtual Environment
Abstract
Recently, there have been many breakthroughs in computer vision thanks to Deep Learning. The availability of huge annotated datasets was one of the reasons of this success, if not the most important. However, these datasets have usually been static datasets, with little to no possibility of interaction with the learning agent. Interaction with the environment is critical if one desires to learn in a human-like fashion. Since placing a learning agent in a real environment could be dangerous to things and people, a promising solution is to use Virtual Environments: 3D graphics engines that can simulate real-like scenarios, allowing some degrees of interaction with the environment. In this seminar, we review some of the available Virtual Environments: RAGE, the graphics engine for videogames such as GTA; Habitat AI, a large-scale scenarios Virtual Environment developed by Facebook AI; AI2-Thor, a Virtual Environment with interaction at its core; finally, we explore the pros and cons of building a custom Virtual Environment.