Széchenyi Plusz RRF

The Technical Sciences Section of the Hungarian Academy of Sciences (HAS) cordially invites you to

Tamás Szirányi’s (corresponding member of the HAS)

inaugural lecture.

 

The title of the talk: ’Machine Perception - Intelligent Space’.

Date: 13 Dec 2022

Time: 11.00 a.m.

Venue: HAS building (1051 Budapest, Széchenyi István tér 1.), Great Hall.

Language: The lecture will be held in Hungarian.

 

The summary of the talk is as follows.

The emergence of "big data" and the "internet of things" is closely followed by the spread of "smart sensor networks" and thus also the spread of the "artificial spatial intelligence", which is a new scientific field within artificial intelligence. The aim of the researchers working in this field is to research computer vision algorithms that enable robots and other intelligent devices to map the 3D spaces around them, to accurately and robustly locate their own position within these spaces, to recognise objects that are present there, and to interpret the movements and interactions of these within the surrounding universe, thereby ensuring the cooperation between robots, intelligent devices, objects and intelligent agents.

In this lecture, the speaker will present methods that help building semantically meaningful spatial models based on visual information of the environment. Many of these spatial models are specified through elegant mathematical descriptions.

The generation of such descriptions are facilitated by the fact that the structure of the visual world can be described by axiomatic properties, e.g., via Markovian graph structure, or via scale-independent image properties. The Fusion Markov Random Field image segmentation and the adaptive texture detection, for instance, are based on and are derived from such properties. The double iteration scheme of blind deconvolution can produce a relative focal depth-map from single-view images relying on a special convergence criterion. Using a similar double iteration scheme, an ergodic regular Markov chain derived from motion correlations can be used to register cameras with large baseline distances.