Page - 82 - in Proceedings - OAGM & ARW Joint Workshop 2016 on "Computer Vision and Robotics“

independent equations for the x- and y-position, thus we need at least 3 correspondences to solve the 6degreesof freedomgivenbyRk andtk. Asmentioned inSection4.3.,outliers (wrongmatches)are possible, thus inpracticeweuseat least4correspondences. In a first step, we compute the reprojection error of all correspondencesC (4) using the initialized camera pose and the mean reprojection error C¯. Then we take all correspondences with an error smaller than 1.5C¯ or a threshold C. We then optimize with the same cost function (4) as for the initial bundle adjustment with the major difference that only the pose of the newly added camera k is optimized, while the rest of the systemMcurr is fixed. After optimization, we again filter bad correspondences with the same approach as described above and perform a second optimization, which is usually very fast due to the already good estimation. We iterate through all images until no morecanbeadded, i.e. donot fulfill anyof theconditions. GlobalBundleAdjustment While keeping the camera systemMcurr fixed and only optimizing the new camera k is very fast and gives an estimate of the model structure, it does not replace a global optimization approach. We perform a final global bundle adjustment step, where all the camera poses are optimized. In this case, we take all the correspondences used during the iterative bundle adjustment step and initialize the camera poses with the previously computed rotation and translation. Here, we also use the cost function presented in (4). Similar to the iterative bundle adjustment, we again filter outliers with the reprojection error after optimization, but instead impose that the error must be smaller than< 1.2C¯. After thisfinal filteringstep, weperformone lastglobalbundleadjustment. 5. ResultsandDiscussion In this section, we present early results of our guided sparse camera pose estimation. We evaluate our approach on two datasets from different buildings, one with a well-textured facade and one with redundant structures. As reference, we use the open-source SfM framework OpenMVG, which can achieveanaccuracyofaround1cmin idealcases [10]. ItutilizesSIFTfeaturesandmanycorrespon- dences to estimate camera poses and a point cloud. OpenMVG chooses the starting views randomly and in our evaluation we had to start SfM multiple times to get a reconstruction. We evaluate the distance between the camera centers generated by the two approaches. OpenMVG estimates the re- construction up to scale, thus to metrically measure the distance between cameras, we transform its worldcoordinate systemto oursusinga robust similarity transform. Figure3 showsahistogram, where the bins showthenumber of distancesbetweencamera centers in therespectiverangeincm. Camerasinthefirst fewbinsarecloser toOpenMVG,while thecamerasin the lastbinare farther away. Especially in thefirst experimentweachieve reasonable results and that withonly90correspondencescompared toOpenMVG’s2969,which isa reductionbyafactorof30. In the second experiment, we only use 56 correspondences compared to 1880 in the reference. The histograms show that we are centimeters away from OpenMVGs reconstruction even though we still achievevisuallyappealingresultswhenreprojecting the laserpoints into the images (seeFig.4). Due to the sparsecorrespondences, evenoneunfiltered outlier can decrease thefinal result significantly. 82