3.4. Accuracy assessment
After the bundle adjustment, the residuals of the GPCs were used to evaluate accuracy of absolute orientation and additionally the stochastic model of the results. Here, all direction XL335 were considered separately and the root mean square errors (RMSEs) were calculated. To assess the final positional accuracy of the geometric model, checkpoint coordinates were located in the orthophoto and DEM within ArcGIS2 environment. Because the terrestrial data mainly shows the elevation of the steep gully walls, accuracy assessment cannot be done with 2.5D data. In this regard, checkpoint coordinates were measured within 3D point cloud. All checkpoint deviations were calculated in dependence on direction (X/Y/Z).
Furthermore 3D point distances were analysed to examine the effects of different referencing methods (GCPs/vGCPs) and to verify results of ICP registration. Thereby, the UAV point cloud was used as reference. Due to significant data gaps in the UAV point cloud, the calculation of cloud-to-cloud distances would lead to erroneous results. For this reason, the Poisson Surface Reconstruction (Kazhdan et al., 2006) was used to derive a meshed 3D model of the UAV point cloud. Thus, gaps were interpolated and finally closed. To investigate the positional accuracy, the point-to-mesh distance between the TERRA point cloud and the UAV mesh could be evaluated directionally. That means positive and negative distances are estimated using the normal vectors of the mesh. So far, this approach has been rarely used, e.g. Koutsoudis et al. (2014) examined the accuracy of 3D-photoreconstruction of a Greek statue based on the cloud-to-mesh distance.