Towards Transparent Urban Perception: A Concept-Driven Framework with Visual Foundation Models

Understanding urban visual perception plays a vital role in modeling how people cognitively and emotionally respond to the built environment. However, conventional survey-based methods are limited in scalability and spatial generalization. To address this, we present a transparent and interpretable framework that leverages recent advances in Visual Foundation Models (VFMs) and concept-based reasoning. Our approach, UP-CBM, constructs a task-specific concept vocabulary using GPT-4o and processes urban scene images via a multi-scale visual prompting strategy. This strategy generates CLIP-based similarity maps that supervise the learning of an interpretable bottleneck layer, enabling transparent reasoning between raw visual inputs and perceptual outcomes. Through comprehensive experiments on Place Pulse 2.0 (+0.041 in comparison accuracy, +0.029 in R2) and VRVWPR (+0.018 in classification accuracy), UP-CBM demonstrates superior predictive performance and transparency. These results underscore the value of combining VFMs with structured concept pipelines for robust and scalable urban visual data processing.