Automatic segmentation and architectural space recognition from complex drawings using deep learning

Abstract

While Building Information Modeling (BIM) is becoming a mandatory standard in the construction industry, the automated conversion of legacy 2D technical drawings into structured 3D BIM models remains a critical bottleneck. A primary challenge in this 2D-to-BIM pipeline is the initial semantic extraction: translating unstructured, cluttered graphical data into clean, topologically accurate pixel-level representations. Traditional automated segmentation methods frequently fail to maintain strict geometric accuracy, producing fragmented physical boundaries and topological errors that hinder subsequent vectorization and 3D reconstruction
pipelines. To address these extraction challenges, this research proposes an improved solution based on a multitask semantic segmentation architecture to simultaneously recognize physical boundaries and spatial functions. To overcome the limitations of traditional neural networks in maintaining geometric accuracy, we introduce an Edge-aware Loss function based on the Sobel operator. This function acts as a spatial regularizer, forcing the model to produce sharp boundaries that strictly adhere to the drawing's topological structure. Furthermore, the study experimentally evaluates the replacement of various feature extraction backbones, including VGG16, ResNet, and ConvNeXt. Experimental results on the standard CubiCasa5K dataset show that the model using the ConvNeXt-tiny backbone combined with Edge-aware Loss achieves superior performance with an mIoU score of 78.57% for the boundary task and for the room space task of 72.27%. Compared to the original VGG16-based architecture, the proposed configuration improves Boundary mIoU by 9.8 percentage points (from 68.78% to 78.57%) and Room mIoU by 17.6 percentage points (from 54.70% to 72.27%). These results confirm the practical value of the method for producing high-quality segmentation maps that can serve as input for vectorization and automatic reconstruction of building models.