University of Michigan Researchers Develop Vision-Language Models to Address Urban Blight in Detroit
Researchers from the University of Michigan have introduced a scalable framework using vision-language models to assess urban blight in Detroit. This approach aims to overcome the challenges of traditional residential blight surveys, which are labor-intensive and difficult to maintain at scale. The study utilizes open-source large vision-language models to evaluate housing attributes such as roof integrity, wall damage, and broken or boarded openings. By comparing professional human annotations with model assessments, the researchers found that ensemble learners, which combine multiple models, outperform individual base models in accuracy and robustness. This method provides a cost-effective way to track and manage housing stock conditions, offering a regularly updatable complement to traditional surveys. 
