Wall-Layer Segmentation in Floor Plans for Energy Retrofit Planning

Reconstructing floor plans from 2D construction drawings can help with understanding the energy demands of existing buildings, which is a key step in designing energy-efficient heating solutions for retrofitting. Current floor plan recognition methods treat walls as monolithic entities and do not account for separate material layers. Their applicability for detailed energy modeling is therefore somewhat limited. This paper proposes an image processing pipeline that takes floor plans as input and segments wall layers in addition to the overall geometry. We demonstrate the feasibility of this method using a U-Net-based model trained on both real and synthetic data, highlighting its potential to improve wall segmentation accuracy. This layer-specific analysis offers estimates of wall material composition alongside their geometry. However, the limited availability of annotated wall layer data and the degree of realism in the synthetic data need improvement. Future works should focus on acquiring high-quality real-world data to improve model performance.