Advanced approaches to object recognition and 3D model construction from heterogeneous data

BASIC HALL, October 17, 14:30 — 15:00

Every field of human activity is undergoing a Data revolution. Old business models are reformulated as data science and machine learning problems. Rising complexity of models and demand for insights is not satisfied with 2D images anymore. In the near future most of machine perception systems will become compliant with 3D processing. Analysis of human movement, faces, MRI, CT and other biomedical images, shelf placement in retail, remote sensing data and much more, all require advanced 3D data analysis capability.

There is several ways to attack this problem, some require different data representation formats and others involve using more complicated sensors. New machine learning models enable us to extract 3D information from 2D videos, laser and infrared scanners provide data in form of Point Clouds, even wi-fi signals are carrying information about people in sparse form.

However, the benefits of 3D data do not come without storage and computational costs, and we try to develop methods to minimize those costs.

In this talk I will underline some applied problems of 3D data analysis, some new algorithms and data structures that help extract information without redundancies and noise. Also I'm going to describe new approaches to 3D machine perception which require strong priors for objects and environments, and special architectures of deep neural networks for Point Cloud processing, and their strengths and weaknesses.