Vol. 45, Issue 4, pp. 559-571

Abstract

This article concerns a key topic in the field of visual object recognition – the use of features. Object recognition algorithms typically rely on a fixed vector of pre-selected features extracted from 2D or 3D scenes, which are then analyzed with various classification techniques. On the other hand, the activation of particular features in biological vision systems is hierarchical and data-driven. To achieve a deeper understanding of the subject, we have introduced several mathematical tools to estimate multiple RGB-D features’ relevance for different object recognition tasks and conducted statistical experiments involving our database of high quality 3D point clouds. From the thorough analysis of the obtained results we draw conclusions that may be useful to design better, more adaptive object recognition algorithms.