6.3. Summary 93The proposed Device Adaptation technique shows a promising increase of accuracyfor the Iconia device but only inconclusive results for the Nexus. The localizationexperiments, that were conducted over both devices, lead to the conclusion thatthe trained propagation models are able to generalize over multiple devices. Thesystem reports similar error rates if the propagation model is trained with eitheronly measurements from the Iconia or from the combined measurements of bothdevices.The final evaluation step, the investigation of how much granularity of the 3D geom-etry is needed for the PHOTON generated propagation models that are subsequentlyused for estimating the localization errors, lead to an interesting result. The bestperforming radio propagation variant is driven by a 3D geometry only consisting ofconcrete material, light walls and three different types of doors(iron, glass, normal).The full detail model that contains additional elements like windows and furnitureis somewhat overspecified for the PHOTON localization algorithm pipeline. Butthis leads also to the more positive result, that sufficient detailed geometry can prob-ably be generated automatically from rasterized 2D floor maps, since both types ofwalls and the location of doors can be obtained from them in most cases.