Residential building type classification from street-view imagery with convolutional neural networks

Residential building type classification from street-view imagery with convolutional neural networks

연구 분야: Artificial Intelligence

학회: Signal, Image and Video Processing

Computer vision techniques are increasingly used to develop efficient and automatic methods that provide alternative data sources. Micro-level information on buildings and physical infrastructure is very important to social, economic, and environmental statistical programs. In this paper, we present how multiple convolutional neural networks (CNNs) are finetuned to classify residential buildings into five different types (e.g., single-detached, semi-detached, condominium apartment, etc.) from their street-view images. A framework for collection and automatic preliminary labeling of street-view images is developed and presented. Furthermore, a big dataset of street-view images of residential buildings has been compiled and labeled to be used for training. Multiple state-of-the-art CNNs are finetuned to accomplish the classification task. The trained models provide a proof of concept and show that CNNs can be used to classify residential buildings using their street-view imagery. The performance of the trained CNNs is measured and presented. This approach can be used to augment the information available on openly accessible databases, such as the Open Database of Buildings.