BIM based building designs include the use of geometric and semantic information for their elemental tasks in the construction industry. The Level of Development (LOD) is widely used in BIM to specify which information must be available at what time during the entire phase of a construction process. LOD helps define the level of maturity and detailing at an instance of the construction process and is considered legally binding information for various evaluations. Many open tools in BIM and commercial software are available that can provide automatic validation of the semantic information of a building model. But automatic validation of the required geometric information needed for a model to fulfil its purpose is still unexplored. Currently, geometric validation is done based on human experience and it still remains a manual task. This thesis study presents a deep learning framework that can automatically evaluate and detect the level of Geometry (LOG) of building elements. The study initially analyses the effectiveness of popular methods available in deep learning for the classification of 3D models for its LOG (e.g., Mesh CNN, Graph CNN, Triple Input CNN, Multiview CNN etc.). The feature patterns that represent the LOG levels of the building models were automatically extracted from the visual representations without manual intervention.MVCNN model architecture is further explored for its effective use in a practical application of LOG classification through adaptation of its model architecture as well as different types of training datasets. Multiplane representations of the MVCNN showed that they were able to classify building elements to different LOG levels with an accuracy of 83%. For commercial applications, a future framework is proposed that has improved prospects in recognising all the feature patterns of different kinds of building elements.
«
Login
BibTeX