Street-Level Facade Understanding with Deep Learning

Description

Urban building databases such as OpenStreetMap contain rich geometric data but lack semantic facade attributes - material, color, age period, and ground-floor use - which are important inputs for urban planning applications. This project develops an automated pipeline that combines street-level imagery (e.g. Mapillary) with neural network-based classification to infer missing facade attributes at city scale, enriching sparse crowd-sourced datasets with automatically predicted building properties.

Tasks

• Query street-level imagery APIs to retrieve georeferenced facade images matched to building footprints via spatial join 
• Train or fine-tune a CNN classifier (e.g. ResNet, EfficientNet) to predict facade attributes such as material, color, and construction period from raw imagery
• Validate predictions against existing ground truth tags in OSM or similar databases
• Export enriched building attributes as GeoJSON for downstream use in urban analytics applications

Requirements

• Knowledge of English language (source code comments and final report should be in English)
• Knowledge of neural networks and deep learning frameworks (e.g. PyTorch, TensorFlow) is advantageous
• Knowledge of web-based APIs and web development is advantageous
• More knowledge is always advantageous

Environment

The project should be implemented as a standalone application, desktop or web-based (to be discussed).