Alessio Del Bue, Marco Crocco
The tutorial will review state-of-art techniques aimed at solving the problem of 3D room reconstruction from sound, i.e. the inference of 3D position of the planar boundaries of an enclosure, from a set of audio signals acquired by one or more microphones, given one or more acoustic sources emitting a signal.
3D room reconstruction from sound is closely related to a set of problems, being currently subject of active investigation, such as source localization and tracking, microphone self-calibration, and signal dereverberation. The tutorial will explain in detail how techniques previously devised for solving such problems can be exploited and adapted to the room reconstruction problem, and, in turn, how knowledge of 3D geometry may improve the solution of the above problems. Depending on the amount of prior knowledge on the microphone and source positions, the emitted signal and the geometry of the enclosure (number of walls, convexity etc.), a number of sub-problems can be defined, from the most controlled scenario to the completely unconstrained problem “in the wild” and the tutorial will organize such sub-problems into a comprehensive taxonomy.
Finally, in order to encourage comparative evaluations of current and future methods, a new dataset based on audio measurements taken in real rooms will be provided and an evaluation protocol, relying on ground truth geometric data will be described. A discussion on open problems and future research directions will conclude the tutorial.