UrbanPhysicalDisorder-4K: Understanding Urban Perception via Counterfactuals and Street View Signs of Physical Disorder

This paper presents a novel framework for explainable urban safety perception analysis using counterfactual reasoning and human-readable interpretations. We leverage a collection of 3,659 street-level images annotated with perceptual safety scores. Unlike traditional segmentation approaches that return only general scene categories, we enrich the visual data with custom manual annotations of urban physical disorder (UrbanPD) elements (e.g., graffiti, broken infrastructure, overhead cables). Our goal is to classify safety perception from urban imagery and understand the causal impact of specific visual elements. To achieve this, we generate counterfactuals by adding or removing disorder-related elements within the scenes. Rather than relying on vector-based explanations, we translate these counterfactual edits into natural language using a large language model, yielding intuitive insights into how specific elements influence safety perception. Our findings indicate that a subset of disorder elements—particularly overhead cables, garbage, and structural damage—has the greatest impact on the perception of unsafe streets.