Topological Semantic Mapping by Consolidation of Deep Visual Features

Many works in the recent literature introduce semantic mapping methods that use CNNs (Convolutional Neural Networks) to recognize properties images. The types of (eg.: room size, place category, and objects) their classes kitchen bathroom, for category) are usually predefined restricted a specific task. Thus, all visual data acquired processed during construction maps lost only recognized remain on maps. In contrast, this work introduces topological method uses deep features extracted by CNN (GoogLeNet), from 2D images captured multiple views environment as robot operates, create, through averages, consolidated representations regions covered each node. These allow flexible recognition other tasks. Experiments with real-world indoor dataset showed is able consolidate them objects categories properties, indicate location images, very promising results.