In this experiment we focused on comparing the estimated geometry with the ground truth data available. For measuring the accuracy of the reconstruction we chose to compute the pixelwise differences of a set of corresponding depth maps. One depth map of each correspondence pair belongs to the input image and was generated from the ground truth model, the other one was estimated during the reconstruction process. Ideally, the pixelwise differences should be zero, because both models share the same coordinate frame. Figure 2 shows an exemplary depth map pair and its pixelwise
difference. The mean depth error was 0.0263 [m] with a standard deviation of 0.0297 [m] for eight depth map pairs. Given the maximum extent of 4 m, the error is about 0.6575% relative to the trench’s size. As figure 2 shows, most of the errors occur
at discontinuities. We hope to further minimize these by an improved fusion of the
depth maps during the reconstruction process.