3. Estimate the camera poses. In contrast to [11], we use the epipolar geometry only for the first two cameras of the sequence. Since we assume the camera calibration to be known, we perform the initialisation of the first two cameras using the essential matrix [10]. This allows the re construction to be performed metrically, instead of projectively. The poses of the remaining cameras are then estimated using the POSIT algorithm [3], so that all poses are reconstructed within the same coordinate frame. If the camera was moved in an orbit around the scene, we use the LoopClosing algorithm from [17] to distribute the drift error amongst all cameras. The pose estimation is concluded by a global bundle adjustment [15].
The reconstructed camera path is shown in figure 1.