Building a balanced and well-rounded dataset for railway asset detection

The entire railway network in Europe has a total length of close to 200 thousand kilometres and is one of the main components of European infrastructure (Eurostat Database 2021). Modernising and maintenance is a sizable effort, and due to the long lifespan of railway links, documentation is discontinued, incomplete, or lost. Using survey methods and recreating accurate as-is documentation improve the efficiency and effectivity of maintaining the rail network. In this paper, we present one major building block in creating such a recognition model. While focusing on images and semantic segmentation, the paper describes how a well-rounded dataset for training ML models can be constructed efficiently. Such a dataset is the missing part in adapting modern image recognition systems to railways and providing semantic information for a fully usable building information model (BIM).