Answers are grounded in this paper and cite their sources.
Street view analytics (SVA) has gained increasing value in environmental criminology. However, most of the analyses are based only on limited factual information, which could lead to potential bias and limit opportunities to investigate the effect of a specific factor of interest while holding others constant. Generating counterfactual street views using vision language models (VLMs) provides a promising avenue for counterfactual reasoning and allows for the controlled study of individual factors. This study presents a novel approach that generates diverse counterfactual street views and compares them with their factual counterparts, using illegal dumping sites in urban environments as a case study. The baseline is a dataset of 757 Google Street Views (GSVs), of which 506 are illegal dumping sites and 251 are controls. Seven VLMs were benchmarked, and Nano Banana Pro was chosen to generate 1000 more counterfactual views with different featured factors. Of these, 921 passed quality screening and were retained for the main analysis. The data was analyzed by two high-powered VLMs namely GPT-5.5 and Claude Sonnet 4.6. In both evaluator graphs, visible litter and surface defects were associated with higher model scores for identifying illegal dumping sites, and well-maintained pavement with lower scores. The resulting graphs describe responses to controlled visual edits and provide hypotheses about environmental influences for future field validation. This approach can be scaled to broader studies of location-based environmental crime.
No takes yet. Share an insight, caveat, or question.
Zhang et al. (2026) studied this question.
Synapse has enriched 10 related papers. Consider them for comparative context: