Many homes in Detroit are in poor condition and may not be safe or livable. These issues, such as broken windows, damaged roofs, or boarded doors, affect the health and safety of the people who live nearby. Traditionally, city workers survey these homes by hand, but this process takes a lot of time and money. Our study looks at how computer models can help identify homes that may be uninhabitable by analyzing images taken from the street. We tested four different computer models that can “see” and “describe” pictures, similar to how people might describe what they see. These models looked at images of houses in Detroit from different angles and were asked to identify signs of damage. We then compared the models’ results with human evaluations.

We found that some models were better than others at identifying certain types of damage. Models that used images from multiple angles gave more accurate results. One model, Qwen2.5-VL, matched human judgments the most closely. This research shows that these tools could one day help cities quickly find and fix dangerous housing conditions, improving public health and safety.