Room usage semantics in models of large indoor environments such as public buildings and business complex are critical in many practical applications, such as health and safety regulations, compliance, and emergency response. Existing models such as IndoorGML have very limited semantic information at room level, and it remains difficult to capture semantic knowledge of rooms in an efficient way. In this paper, we formulate the task of generating rooms usage semantics as a special case of room classification problems. Although methods for room classification tasks have been developed in the field of social robotics studies and indoor maps, they do not deal with room usage and occupancy aspects of semantics, and they ignore the value of furniture objects in understanding room usage. We propose a method for generating room usage semantics based on the spatial configuration of room objects (e.g., furniture, walls, windows, doors). This method uses deep learning architecture to support a room usage classifier that can learn spatial configuration features directly from semantically labelled point cloud (SLPC) data that represent room scenes with furniture objects in place. We experimentally assessed the capacity of our method in classifying rooms in office buildings using the Stanford 3D (S3DIS) dataset. The results showed that our method was able to achieve an overall accuracy of 91% on top-level room categories (e.g., offices, conference rooms, lounges, storage) and above 97% accuracy in recognizing offices and conference rooms. We further show that our classifier can distinguish fine-grained categories of of offices and conference rooms such as shared offices, single-occupancy offices, large conference rooms, and small conference rooms, with comparable intelligence to human coders. In general, our method performs better on rooms with a richer variety of objects than on rooms with few or no furniture objects.