Creating Consistent Scene Graphs Using a Probabilistic Grammar

摘要

Growing numbers of 3D scenes in online repositories provide newrnopportunities for data-driven scene understanding, editing, and synthesis.rnDespite the plethora of data now available online, most ofrnit cannot be effectively used for data-driven applications because itrnlacks consistent segmentations, category labels, and/or functionalrngroupings required for co-analysis. In this paper, we develop algorithmsrnthat infer such information via parsing with a probabilisticrngrammar learned from examples. First, given a collection of scenerngraphs with consistent hierarchies and labels, we train a probabilisticrnhierarchical grammar to represent the distributions of shapes,rncardinalities, and spatial relationships of semantic objects withinrnthe collection. Then, we use the learned grammar to parse newrnscenes to assign them segmentations, labels, and hierarchies consistentrnwith the collection. During experiments with these algorithms,rnwe find that: they work effectively for scene graphs for indoorrnscenes commonly found online (bedrooms, classrooms, and libraries);rnthey outperform alternative approaches that consider onlyrnshape similarities and/or spatial relationships without hierarchy;rnthey require relatively small sets of training data; they are robustrnto moderate over-segmentation in the inputs; and, they can robustlyrntransfer labels from one data set to another. As a result, the proposedrnalgorithms can be used to provide consistent hierarchies forrnlarge collections of scenes within the same semantic class.