How Many Visual Levers Drive Urban Perception? Interventional Counterfactuals via Multiple Localised Edits
Researchers have developed a new framework to identify specific visual elements that influence human perception of urban scenes, moving beyond simple correlations. This interventional counterfactual approach systematically tests how localized edits to images, such as changes in mobility infrastructure or physical maintenance, can alter predicted safety judgments. The framework aims to provide a more robust understanding of scene explainability by generating and validating counterfactual edits based on realism and plausibility, with human judgment serving as the ultimate validation. AI
IMPACT Introduces a novel method for understanding how specific visual changes impact AI perception of urban environments, potentially improving model interpretability.