FRAMER: Fast Reconstruction of Architectural Models from Existing Resources

FTL Labs Corporation, together with collaborators at the National Institute of Standards and Technology and University of Massachusetts professor Dr. Evangelos Kalogerakis, proposes a powerful new system for automatic 3D building model synthesis. FTL's “FRAMER” (Fast Reconstruction of Architectural Models from Existing Resources) system is designed to be flexible, fast, and accurate. It combines specially trained neural networks with advanced 3D processing algorithms and intelligent data exports to enable high quality building model generation from both blueprints and photographs, and parameter exports for accurate transport and dispersion simulation. FTL will leverage its ongoing research utilizing neural networks for image and 3D data processing, object detection, and high-quality 3D model generation for rapid development. This will enable early demonstration of a software system that provides the highest fidelity 3D building models based on the data used as input and the most accurate possible exports to AR and VR applications, in addition to T&D software such as NIST’s CONTAM. Blueprints constitute a unique problem for neural networks due to their widely varying range of quality, frequent lack of relevant information, and ambiguous distinctions between room types. FTL overcomes these challenges with a unique synthetic data training step that leverages Dr. Kalogerakis’s research with a large dataset of accurate and annotated building models proven to increase neural network accuracy for building part and object detection. This data, which will be extended to include accurate blueprint output, enables new and existing neural networks to be trained easily and repeatedly, increasing the robustness of detection for typical objects such as walls, doors, and windows. The result of this NN-processed blueprint is a data structure containing all the building’s relevant features, from which a 3D building model can be procedurally generated. FRAMER’s generated 3D models will optionally be segmented and include objects labeled using state of the art neural network research, bringing a richer experience to the existing virtual and augmented reality applications in use at CBD. These key developments also include the use of an additional neural network to automatically augment a 3D indoor scene with new objects and furnishings that match their surroundings. This exciting research will enable FRAMER to provide true-to-life indoor building areas even when photos or scans of those rooms do not exist. Through FTL’s collaboration with the developers of CONTAM at NIST, FRAMER’s exported building parameters will support high quality transport and dispersion modeling. The exported building data will be usable directly in CONTAM through the creation of building templates and automatic editing of project files. Additionally, FTL will leverage its extensive experience with the development of AR and VR applications for FRAMER’s high fidelity 3D building model exports.