Instance Segmentation of Fire Safety Equipment Using Mask R-CNN

In the construction field, Building Information Modeling (BIM) offers various application scenarios in the design, construction, and maintenance phases of a building’s lifecycle. The utilization of BIM in the Operation and Maintenance (O&M) phase of existing buildings is a particular challenge, as suitable BIM models are usually not available as a basis. In this context, many researchers have investigated various methods to automatically create BIM models from existing information and data. Focusing on images, numerous methods have been investigated for detecting and classifying certain objects in buildings. However, despite the importance of fire protection in buildings, the research field has not focused on the benefits of the automatic recognition of fire safety equipment (FSE, e.g., fire blankets) in much detail. Particularly in existing buildings, the recurring inspection and maintenance of fire safety equipment is a responsible task and required by law. It is the responsibility of the owners and facility managers to ensure the availability and proper functioning of fire safety equipment in the building. Consequently, this work aims to contribute to this research area by investigating the state-of-the-art Mask Region-Based Convolutional Neural Network (Mask R-CNN) for instance segmentation and object detection of FSE in RGB images. The results show that this approach automatically extracts valuable semantic information that provides the presence of fire safety equipment in images. In addition, this study investigates the influence of hyperparameter adjustment on the detection of FSE objects in indoor scenes. It is also examined how the additional use of augmented data improves the performance of the neural network.