Semantic indoor scenes recognition based on visual saliency and part-based features

Abstract

This paper presents a semantic indoor scene recognition method used for an autonomous mobile robot. The proposed method comprises feature description using accelerated KAZE (AKAZE), saliency maps (SMs) for feature selection, creating bags of visual words (BoVWs) using self-organizing maps (SOMs), and incorporating scene recognition based on category maps using counter propagation networks (CPNs). Saliency-based features are used in semantic indoor scene recognition. This study was conducted to evaluate the combination of salient features. We conducted evaluation experiments using a public benchmark dataset for comparison of feature sets of three types. We demonstrated basic properties of feature combination using part-based key-point feature descriptors according to saliency local regions consisted of generic objects.