SewerOD: A visual sewer disease detection dataset for machine learning

Abstract

The underground sewer pipeline is an essential urban infrastructure that undertakes the vital responsibility of discharging sewage. The location and type of diseases in the pipeline often need to be checked manually by inspectors, which cannot be completed efficiently due to labor costs and time requirements. With the development of computer vision, the use of detection technology to maintain sewer pipelines has extremely high research value. However, image data of pipelines are often regarded as commercial secrets, and these studies are greatly limited due to the scarcity of open-source pipeline disease datasets. To solve this problem, we present a public large-scale object detection dataset for sewer disease detection named SewerOD in this work. The dataset contains about 47K images, annotated by professional researchers, and includes two of the most widespread structural disease types: Corrosion and Crack. Our dataset is available at https://github.com/SewerOD.