| Summary: | When integrating railway constructions and refurbishments into an existing infrastructure, it is beneficial to have knowledge of the exact state, geometry, and placement of the connected assets. While new constructions and the maintenance of existing lines can directly use existing digital models and incorporate them into their processes, existing railways often predate digital technologies. This gap in digital models leaves the planning processes of new constructions and refurbishments to primarily rely on non-automated and analogue workflows. With a multitude of asset types, layouts and country-specific standards, the automatic generation of adequate detection models is complicated and needs to be tailored to the current project environment, generating considerable overhead. Addressing this issue, this paper presents the concept of priming. Priming increases the adaptation performance to highly volatile, low-data environments by leveraging previous, existing CAD projects. We introduce a translation scheme that converts the existing 3D models into realistic, project-specific, synthetic surveys and a complemental dialled-in training routine. When applied to a convolutional neural network, we show that the primed training will converge faster and with greater stability, especially when using sparse training data. Our experiments show that priming can reduce the time for network adaptation by over 50%, while also improving resilience to underrepresented object types.
|
|---|