IPD: An Industrial Pipeline Dataset for Anomaly Detection and Localization

This paper examines the challenging task of visual inspection on industrial pipelines. Key challenges in this area include: i) highly noisy and cluttered pipeline background, ii) damage types are not known a-priori, iii) high variability in damage visual appearance, and iv) imbalanced distribution of the training samples (images of damaged pipelines appear less frequently than non-damaged). Therefore, visual Anomaly Detection (AD) and Anomaly Localization (AL) are valid and suitable methods to address such a task. This paper introduces the Industrial Pipeline Dataset (IPD), that is mainly collected from industrial facilities in Europe and focuses on damages to insulated pipelines. Data collection and annotations was guided by consultations with industry experts to ensure the accurate identification of insulation damages. State-of-the-art AD and AL methods were evaluated to establish a baseline for the inspection task. Their rather mediocre performance, shown both visually and empirically, showcases the complexity of the task. As such, this paper makes two main contributions: i) IPD is, to the best of our knowledge, the first publicly available AD/AL dataset on visual inspection of industrial pipeline insulation, making it highly valuable for real-world industrial applications, and ii) since modern AD/AL methods fail to perform well on IPD, it constitutes a valuable research benchmark.