GET-Forschungsseminar Abstracts

An Evaluation Framework and Scoring System for Artificial Visual Attention Models

Adarsh Thirukappanavar Math, GET Lab

Vortrag: Mi. 04.02.2015 16:30, Raum P 1.4.17

Abstract:

The human visual system has the capability to identify relevant parts of a scene at early stages and guide attention to their locations. This ability of shifting attention helps our brain to focus processing on relevant objects rather than on irrelevant ones. If this concept is applied to real-time computer-vision systems, such as mobile robots, the processing power required can be greatly reduced. In order to achieve this feature, the most salient location of an image has to be detected. Hence, for achieving this, various artificial attention models have been proposed which mimic biological visual attention. Even though there are certain methods to evaluate these models, it is difficult to obtain generalizable results, as the outcome often depends on the scene context and the tasks being performed. Therefore, in this thesis a framework which automatically evaluates attention mod- els and creates a combined score that includes different tasks and measures them is proposed. Furthermore, through the output of such a framework one can decide on choosing an appropriate attentional model for a desired application.