Analysing the Potential Impact of Labeling Disagreements for Engineering Sensor Data.

We present the hyperbola recognition problem in Ground Penetrating Radar – GPR – data as an example for pattern recognition in complex engineering sensor data. Traditionally, GPR data are analyzed manually by human experts in a tedious and time-consuming process, e.g., to deduce the positioning of linear object underneath roads just before reconstruction works take place. For supporting this process using Machine Learning methods, one needs to have accurate ground truth data to derive models out of it. As an accurate acquisition of such annotated data is impossible even for a quasi-ideal case, we annotated 700 radargram images manually. This paper presents and discusses the outcomes of this study and concludes, that using just a single evaluation criteria to compare performances of GPR-focused Machine Learning methods might not be enough.