IncivilityCaps: A Multimodal Dataset for Image Captioning in Urban Incivility Scenes

Urban scene understanding, a critical area in computer vision and urban computing, focuses on interpreting urban environments. However, the lack of domain-specific datasets for urban incivility scenes, such as littering and vandalism, hinders the development of effective detection models. To address this gap, we introduce IncivilityCaps, a comprehensive dataset comprising 40,000 images depicting urban incivility. Each image is annotated with detailed descriptions, classified into one of 44 categories of incivility, evaluated against 6 standardized inspection criteria, and associated with a date. We provide a comprehensive analysis of the dataset and establish benchmark experiments using the OFA model. This dataset holds significant social value by facilitating the development of automated systems to detect and address urban incivility, thereby promoting cleaner and safer urban environments. Furthermore, to promote further research and collaboration, the IncivilityCaps dataset is openly available on Hugging Face.