A Self-Verification Evaluation Method of Accuracy for RGBD 3D Reconstruction

3D reconstruction with RGBD images is a popular research issues in 3D vision as the easily accessibility of depth camera which can out depth images directly. The accuracy evaluation of a 3D reconstruction algorithm is indispensable no matter for academic research or for industrial application. Most evaluation methods used now, such as Absolutely Trajectory Error or 3D surface error, rely on the ground truth values provided beforehand as reference. These methods can only be used on datasets providing these reference values which are difficult to obtain. Actually for the practical reconstruction of a physical object or scene, there is no ground truth to refer and it is hard to assess the quality of the reconstructed results. A self-verification evaluation method of accuracy for RGBD 3D reconstruction without need any ground truth is proposed in this work to address this issue. During the reconstruction process, we generate depth images based on the estimated camera poses and 3D model. With the input depth images as reference, an accuracy metric is elaborated by computing the differences between the generated depth values and the inputted depth values. Experimental results illustrate the effectiveness of our proposed method which can give evaluation results consistent with the ATE metric.