We're Building a Wall - Size-Invariant Semantic Segmentation of Floor Plan Images

Sammanfattning

The automation of floor plan interpretation and reconstruction is a key challenge in architectural visualization workflows, particularly when handling heterogeneous, real world data. This thesis project investigates the use of deep learning–based semantic segmentation combined with geometry-aware postprocessing to automate the extraction of structural elements from floor plan images. The work is motivated by an industrial use case at NORNORM, where manual annotation of customer-provided floor plans represents a significant bottleneck in the generation of 3D visualizations. A convolutional neural network based on a UNet architecture with a pretrained ResNet backbone is developed to segment walls, windows, and doors from floor plan images. Special attention is given to dataset-related challenges, including large variations in image scale, annotation quality, class imbalance, and the presence of corrupted data. Multiple preprocessing strategies are evaluated to analyze their impact on model performance and generalization. In addition, the effect of augmenting proprietary data with the public CubiCasa5k dataset is examined. Beyond pixel-wise segmentation, a postprocessing pipeline is introduced to convert raster predictions into structured, vector-based geometric representations suitable for downstream 3D reconstruction. This pipeline integrates skeletonization, graph-based path extraction, and line simplification using the Ramer-Douglas-Peucker algorithm, producing clean and scalable geometric outputs. Model performance is assessed using standard segmentation metrics as well as uncertainty measures to estimate prediction reliability in deployment scenarios without target annotations available. The results demonstrate that models trained on tiled images yields the most robust performance across diverse inputs. The proposed system shows strong potential to significantly reduce manual effort while maintaining high accuracy, supporting scalable and sustainable architectural visualization workflows.