Deep Learning for Building Exterior Cladding Classification Using Pre-trained CNNs

This project applies transfer learning using ResNet50 and InceptionV3 to classify building façade cladding materials from labeled Google SVIs.

Objective

Automatically classify exterior cladding types—such as Brick, Concrete, Curtain-Wall, Mixed, Others, and Stone—to support scalable building stock analysis for energy and urban modeling.

Dataset

• Source: Wang et al., 2024 – Building Façade Dataset
• Cities: London (for training) and Scotland (for generalization testing)
• Format: Image folders with class-wise subdirectories

Models

• ResNet50 and InceptionV3 with frozen base layers and custom dense heads
• Trained separately on unaugmented and augmented datasets

Key Results

Model	Dataset	Test Acc.
ResNet50	Augmented	68.2%
InceptionV3	Augmented	70.4%

Notebooks

• final_project_unaugmented.ipynb
• final_project_augmented.ipynb

Next Steps

• Fine-tune base layers
• Try Vision Transformers (ViT) or Swin Transformers Liu et al., 2025
• Work on domain adaptation.

Requirements

Python 3.9 · TensorFlow · NumPy · Matplotlib · scikit-learn · seaborn
(Optimized for Apple Silicon with tensorflow-metal) - 1 GPU

© 2025 Meltem Sahin Ozkoc – Carnegie Mellon University