CUBIT-Det: High-Definition Multi-Scenario Defect Dataset for Infrastructures Fully Evaluated with Deep Learning Processes

Civil infrastructure defect inspection is a critical yet intricate task. Unmanned robots integrated with the deep learning-based visual inspection are more comprehensive and efficient methods with much fewer safety concerns than manual visual inspection. Despite the success of deep learning in numerous object detection tasks, its potential for infrastructure defect inspection has not been fully explored under the lack of annotated high-quality datasets with large volumes. This paper establishes CUBIT-Det, a high-resolution infrastructure defect dataset comprising 5527 optical images with resolutions up to
$8000 \times 6000$
, obtained from multiple scenarios by using professional acquisition equipment on various robotic platforms. CUBIT-Det covers a broad spectrum of practical situations, backgrounds, and defect types compared to existing defect datasets. Extensive benchmarking experiments among state-of-the-art detection methods have been conducted on the proposed dataset to validate its feasibility. To the best of our knowledge, this study is the first to comprehensively benchmark infrastructure defect detection methods on annotated high-resolution defect images, laying a solid baseline for the future development of defect inspection solutions. The reliability and feasibility of our dataset have been further demonstrated by the real-world infrastructure defect inspection task.