Reproduce “CityPulse: Fine-Grained Assessment of Urban Change with Street View Time Series” on OpenResearch

About this paper

Tianyuan Huang, Zejia Wu, Jiajun Wu, Jackelyn Hwang, Ram Rajagopal

Researchers at Stanford University and UC San Diego developed CityPulse, a framework for fine-grained urban change assessment utilizing the largest street view time series dataset to date and a DINOv2-based deep learning model. The model achieved an F1-score of 87.96% in detecting urban changes and showed statistically significant correlations with shifts in median household income and population size, outperforming traditional proxies like construction permits.

Key insights

• Street view time series data provides a powerful "natural data augmentation" effect, enabling models to robustly distinguish genuine structural changes from noise (e.g., lighting, transient objects) more effectively than artificial augmentation techniques.
• Generically pre-trained vision transformers like DINOv2 are highly effective backbones for street-level change detection, demonstrating robust feature learning capabilities even with limited domain-specific training data.
• Fine-grained urban changes detected from street view imagery offer a more accurate and reliable visual proxy for socio-economic shifts in neighborhoods compared to traditional data sources like construction permits.

Summary by alphaXiv.