With the prevalence of devices equipped with 3D data acquisition technologies (i.e. color and depth sensors), understanding 3D scenes from scanned data was seen as a significant attraction. In addition, deep neural networks (DNNs) have shown remarkable performance in various 3D applications, such as the construction of building information models (BIMs) from digitized 3D data (i.e. from digitization to BIM). However, several problems, such as the lack of labeled data for DNN training, poor extraction of feature descriptors for object recognition and inefficient approaches to parametric reconstruction, call into question the effective use of DNNs for digitization to BIM. To overcome the aforementioned problems, in this research we investigate a two-stage framework for performing BIM-based scene reconstruction from digitized data. In particular, in the first stage, we investigate a semi-supervised object detector with a geometry-aware point cloud (PC) feature extraction backbone. Secondly, we aim to develop an Industry Foundation Class (IFC) object extraction approach to match the object detected in the scanned scene with pre-registered IFC parametric objects. The proposed framework exploits the remarkable performance of DNNs to understand parametric scenes from scanned data, and enables “as-is” BIM reconstruction without the need for traditional labor-intensive methods.