Inference of relevant BIM objects using CNN for visual-input based auto-modeling

This paper aims to propose an approach to inferring relevant BIM objects using techniques for recognition of design attributes and calculation of visual similarity using a deep convolutional neural network. Main objective is a visual-input based modelling approach to the automated building and interior design, and this paper represents a preliminary yet critical part of the process. While designing a building, it is important to consider requirements such strong constraints (e.g. regulations, a request for proposal) as well as relatively weak and qualitative constraints such as preference style of clients or users, design trend and etc. Building Information Modeling (BIM) enables to execute to check and review a building design according to the constraints. Until now there is no research that has focused on relatively qualitative and “soft” constraints such as preference or design trend. As a part of research on design supporting system for such “soft” constraints, this paper focuses on training deep learning models that recognize design attributes of BIM object, calculate visual similarity with other objects, and for visual-input based auto-modeling on BIM using the models. A deep convolutional neural network is utilized to extract a visual feature of the 3D object. The input data type for extracting feature data is a 2D rendering image of an object with a specific view and option. The target object is a chair. The feature data is used as input to training models inferring design attributes such as design style, seating capacity and sub-type of a chair and also calculating the visual similarity between objects. This models plays an important role of visual-input based automated modelling system.