Towards Large Scale Facade Feature Extraction Using Vision Transformers & Panoramic Street View Imagery

Paper in proceeding, 2025

In response to the growing need for detailed semantic representations of built environments, this article introduces an automated and scalable pipeline for facade feature extraction using street-level panoramic imagery. The proposed method tackles the complexities of facade parsing by integrating the state-of-the-art Vision Transformer SegFormer architecture for efficient semantic segmentation, thereby enhancing both accuracy and performance. The workflow spans the entire process: from the automated collection of building footprints and panoramic images, through image preprocessing and correction, to precise segmentation and subsequent analysis of facade elements. Using open-source and easily accessible tools, the approach reduces the barrier to entry, fostering broader adoption in urban analytics, sustainable planning, and infrastructure assessment. Experimental evaluations demonstrate that the method effectively captures both local details and global context, yielding highquality semantic representations of building facades. Overall, the proposed method contributes a comprehensive, end-to-end framework that not only advances automated urban modeling, but also provides a robust and adaptable solution for detailed facade feature extraction in complex urban settings.