*This dataset supports online use.Click here to jump.

ConstructionSite is a multimodal benchmark dataset for construction site scenes released by the University of British Columbia and the University of British Columbia in 2025. The related paper results are "Are Large Pre-trained Vision Language Models Effective Construction Safety Inspectors?", which aims to evaluate and improve the image understanding and reasoning capabilities of vision-language models in construction safety environments.

This dataset contains 10,013 construction site images, divided into a training set of 7,009 images and a test set of 3,004 images. Each data entry includes an image, an image description, a question and answer about safety rule violations, bounding box annotations of the violating object, specific categories for object detection tasks (such as excavators, rebar, and workers wearing white hard hats), and image attributes such as lighting, camera distance, viewing angle, and information quality. This dataset features complex scenes, diverse annotations, and is close to actual construction safety inspections. It is suitable for tasks such as image description, visual question answering, object detection, visual localization, and multimodal reasoning.