Abstract

Accurate building information is essential for urban planning, occupancy estimation, energy demand
analysis, and the development of reliable 3D city models. Among building attributes, the number
of floors is particularly important because of its impact on energy consumption and occupancy.
However, floor-count information is often missing in urban datasets, even when building footprints
are widely available.
This limitation is evident in open building-data platforms such as Colouring Cities, which provides a
harmonized data structure for mapping and visualizing urban building data across cities worldwide.
Motivated by this challenge, this thesis presents a pipeline for estimating building floor count from
facade images using deep learning, including a comparison between U-Net and SegFormer for
window segmentation.
While most studies rely on convolutional neural networks, transformer-based models remain less
explored for floor-count estimation despite their capability to handle facade variations, lighting
conditions, and occlusions. This gap motivates the investigation of transformer-based models within
the proposed pipeline.
The workflow consists of three steps: window segmentation using SegFormer, spatial clustering of
detected windows into vertically aligned groups representing candidate floors, and post-processing
to remove unrealistic floor detections. The final output is one unique floor-count estimate per facade
image, which can be integrated into building datasets.
The results answer two research questions: (1) how accurately floors can be estimated using the
full pipeline; (2) whether SegFormer improves window segmentation compared with U-Net. On the
LSAA dataset, the pipeline achieved approximately 70% exact floor-count accuracy, and SegFormer
improved IoU from 0.742 to 0.777 over the U-Net baseline.