A Highly Accurate 3d Reconstruction method for Small Object Scanning

In this paper, we propose a 3d reconstruction scheme for micrometer-level scanning of decimeter-scale objects, with both real-time efficiency and high accuracy, by improving classical point cloud registration methods and graph-based optimization methods. We find several major factors which significantly influence the accuracy of point cloud registration in such scanning contexts and discard other less relevant factors with the concern of efficiency. We utilize a group of keyframe selecting strategies to make the initial estimation of camera poses attainable in real-time. To eliminate the accumulated modeling error of localization, we perform online local pose-graph optimization and offline global bundle adjustment as a postprocessing step, by incorporating the metric structure of the graph of the real-time registration results. The numerical experiments indicate that the proposed method can achieve the accuracy required by many industrial applications such as dental restoration and mechanical measurements.