Learning Functional Properties of Rooms in Indoor Space from Point Cloud Data: A Deep Learning Approach

Guoray Cai, Yimu Pan

Abstract: This paper presents a method to derive functional labels of rooms from the spatial configuration of room objects detected from 3D point clouds representation. The method was inspired by the intuition that spatial configuration of room objects has intimate link with the intended functional purposes. To explore the possibility of inferring the room usage information from its spatial configuration, we designed and trained a deep learning model to learn the important features of spatial configuration of room scenes and examined the predictive power of the model in inferring room usage. We present an experiment on using the model to to predict room function category on Standford 3D (S3DIS) dataset, and achieved reasonable performance. Analysis of accuracy and confusion rates allows us to draw a number insight on the separability of rooms among top level categories (such as offices, conference rooms, lounge, hallways, and storage rooms). Our findings suggested that our method is promising, with an accuracy of 91.8% on predicting room function categories. Future work should further validate and refine our method using data with more balanced training samples on the range of room types as they become available.