City Forensics: Using Visual Elements to Predict Non-Visual City Attributes.

Area of Science:

• Computer Vision
• Urban Analytics
• Machine Learning

Background:

• Cities possess complex non-visual attributes (e.g., crime, housing prices) that are difficult to measure comprehensively.
• Street-level imagery offers a rich, yet largely untapped, source of visual data about urban environments.

Purpose of the Study:

• To develop an automated method for identifying and validating predictive relationships between urban visual appearance and non-visual attributes.
• To create a scalable system for analyzing large-scale street-level image datasets and associated urban data.

Main Methods:

• Utilizing street-level images and location-based attribute data.
• Identifying discriminative visual elements within images.
• Training non-linear Support Vector Regression models to predict attribute values.
• Implementing a scalable distributed processing framework to accelerate visual element extraction.

Main Results:

• Established predictive relationships between visual elements and various urban attributes, including crime rates, housing prices, population density, and perceptions of danger.
• Demonstrated that the developed predictor significantly outperforms human performance in predicting theft rates from street-level images (33% higher accuracy).
• Validated the system's efficiency and scalability across six American cities.

Conclusions:

• Urban visual appearance is a significant predictor of diverse non-visual city attributes.
• The developed automated system provides a powerful tool for urban data analysis and understanding.
• Potential applications include neighborhood boundary definition, attribute-aware navigation, and visual element validation.