Abstract:
Floorplan understanding is critical for architectural design, reconstruction, and urban planning.
Accurate detection of foundation structures help to determine the stability and the functionality of
buildings. We present a deep learning framework for automated foundationstructure identification
based on a multi-head U-Net with a ResNet encoder, that jointly segmentations foundation-lines
and junction heatmaps. The proposed approach was evaluated on a real-image dataset of 1,000
floorplans annotated with foundation structure lines and junction points. The framework achieved
a precision of 0.9438 and a recall of 0.9484 for junction detection, as well as an intersectionover-
union (IoU) of 0.8806 for foundation-line segmentation. These results outperform other
segmentation models, including DeepLab-v3+ (precision 0.9273, recall 0.9329, IoU 0.8569) and the
vanilla U-Net (precision 0.3731, recall 0.6804, IoU 0.5784), demonstrating robust learning across
diverse layout configurations. Qualitative analysis shows accurate localization of key junctions
and coherent foundation structure predictions across diverse floorplans. This work contributes
a practical pipeline for real-image dataset preparation, a multi-task architecture tailored to
foundation cues, and empirical evidence supporting its effectiveness for structural recognition.
Future work will involve curating and annotating more diverse real-world floorplans to fine-tune the
model and extending the system to explicitly map connections between junctions and foundation
structures for enhanced structural reasoning and downstream engineering applications.
