HOUSED:Housing-dimensiOnal visUal inSpection imagE Dataset

The Housing-dimensiOnal visUal inSpection imagE Dataset (HOUSED) was developed through systematic field investigation and standardized data processing to support research on automated building defect recognition and intelligent urban housing inspection. The dataset was collected in city A (In accordance with the confidentiality agreement, the city in this dataset will be anonymous.), China, between March and August 2024, covering 10 administrative districts, 32 subdistricts, and 484 residential communities. A total of 11,317 residential buildings constructed before 2000 were surveyed, ensuring that the dataset reflects typical characteristics of aging urban housing. Image acquisition was conducted by trained investigators using consumer-grade smartphones and digital cameras under varying lighting conditions, weather environments, and viewing angles, thereby enhancing the diversity and real-world representativeness of the data. A total of 34,051 defect-related images were initially collected. Following systematic data cleaning based on the principles of authenticity and informational validity, 26,351 high-quality images were retained (77.4% of the original data). Images that were severely blurred, heavily occluded, non-informative, or captured from extreme angles were removed. The final dataset contains 74,660 manually annotated defect instances, covering 17 categories and 19 inspection indicators related to building safety and functionality. Each defect instance was labeled with a standardized bounding box and category identifier using YOLO-format annotation tools, with normalized coordinate outputs to ensure compatibility across mainstream object detection frameworks. Annotation quality was controlled through cross-review and consistency checking to minimize subjectivity. To enhance cross-task applicability, a dual-task annotation structure was adopted. In addition to object detection labels, image-level classification samples were automatically generated by cropping bounding box regions with contextual expansion. The classification component follows a single-label structure organized into 17 category folders.