Semantic parsing for priming object detection in RGB-D Scenes

Abstract

The advancements in robot autonomy and capabilities for carrying out more complex tasks in unstructured indoors environments can be greatly enhanced by endowing existing environment models with semantic information. In this paper we describe an approach for semantic parsing of indoors environments into semantic categories of Ground, Structure, Furniture and Props. Instead of striving to categorize all object classes and instances encountered in the environment, this choice of semantic labels separates clearly objects and nonobject categories. We use RGB-D images of indoors environments and formulate the problem of semantic segmentation in the Conditional Random Fields Framework. The appearance and depth information enables us induce the graph structure of the random field, which can be effectively approximated by a tree, and to design robust geometric features, which are informative for separation and characterization of different categories. These two choices notably improve the efficiency and performance of the semantic parsing tasks. We carry out the experiments on a NYU V2 dataset and achieve superior or comparable performance and the fraction of computational cost.