6.2. Localization 91Figure 6.7 The Device Adaptation shows no effect on the error rates of the Nexus.Figure 6.8 Localization errors for radio propagation models trained on materialannotated 3D geometry of different complexities. From left to right, the geome-try becomes more complex. It can be observed that using the most complex ge-ometry of Full11 leads unexpectedly to worse results than using the less complexBasic+Doors5.Under the Iconia trained propagation model, the Device Adaptation for the Nexusshows no significant enhancement. The results become even marginally worse. Un-der the assumption, that the Device Adaptation is a form of final training for radiopropagation model, this result makes no sense. The measurements of the Nexusdevice should have been a valuable information to adapt to the less"familiar" prop-agation model as it has only be trained on the Iconia device.6.2.3.3 Granularity of the 3D geometryIn the following experiment, the impact of the accuracy of the 3D geometry on thelocalization error is analysed. Six radio propagation models were trained on 3Dgeometry that represents the UMIC environment in different levels of complexity.The details of the six models are described in the former evaluation of these modelsin section 6.1.2.1.160 path samples originating from the Iconia device were selected from the evaluationcorpus, as including the samples from the Nexus can lead to undesired device specificdistortions in the outcome of the experiment. The samples are evaluated with allthree localization algorithms which leads to results that are listed in figure 6.8.Unexpectedly, the single material variants Basic1 and Full1 perform worse thanthe other variants. This underperformance was also visible in the reported RPE