Feasibility Study and Evaluation of ML Methods for Automated Condition Assessment of Building Facades

To date, the condition of building facades has been assessed primarily through manual inspections—a process that is time-consuming, subjective, and not feasible on a large scale for extensive building portfolios. Image-based methods offer a potential solution, but they face two challenges: facade damage is visually heterogeneous (cracks, spalling, vegetation growth, discoloration), and there is a lack of labeled data for local building structures. Public facade datasets typically annotate building components rather than damage; available damage datasets primarily originate from different construction and climatic contexts.

The goal of this thesis is to conduct a robust feasibility study: Can the condition of a facade be automatically assessed from an image, which damage classes can be reliably identified, and what are the limitations of this approach? The aim is to achieve a multi-label classification of clearly defined damage features. The thesis may be written in English or German.

Work Packages

1. Literature and Technology Review – Review of existing datasets and published approaches to facade damage detection.
2. Class definition and data collection – Defining a catalog of 3–5 visually distinguishable damage features (e.g., cracks, plaster spalling, vegetation/algae, dirt/graffiti). Building and annotating a custom image dataset.
3. Model Development – Transfer learning on a pre-trained backbone (no training from scratch). Handling a small and imbalanced dataset: data augmentation, class weighting, and appropriate tiling of large facade images.
4. Baselines and Comparison – Comparison of our own fine-tuned model with a vision-language model in zero-shot and few-shot settings, without separate training. (If the VLM performs comparably without training, this constitutes a legitimate and relevant result of the work.)
5. Evaluation and Error Analysis – Evaluation using appropriate metrics; qualitative error analysis: Which types of damage are misclassified, and under what imaging conditions does the model fail (shading, backlighting, obstruction by vegetation or parked vehicles)?

Expected Results

Documented, reproducible code; annotated dataset; evaluation results for the model variants; a well-reasoned statement on feasibility for each damage class, including the limitations of the approach.

Profile

Degree in Computer Science or a related field; knowledge of Python; basic knowledge of machine learning