3DMatch: Learning the Matching of Local 3D Geometry in Range Scans

Establishing correspondences between 3D geometries is essential to a large variety of graphics and vision applications, including 3D reconstruction, localization, and shape matching. Despite significant progress, geometric matching on real-world 3D data is still a challenging task due to the noisy, low-resolution, and incomplete nature of scanning data. These difficulties limit the performance of current state-of-art methods which are typically based on histograms over geometric properties. In this paper, we introduce 3DMatch, a data-driven local feature learner that jointly learns a geometric feature representation and an associated metric function from a large collection of real-world scanning data. We represent 3D geometry using accumulated distance fields around key-point locations. This representation is suited to handle noisy and partial scanning data, and concurrently supports deep learning with convolutional neural networks directly in 3D. To train the networks, we propose a way to automatically generate correspondence labels for deep learning by leveraging existing RGB-D reconstruction algorithms. In our results, we demonstrate that we are able to outperform state-of-the-art approaches by a significant margin. In addition, we show the robustness of our descriptor in a purely geometric sparse bundle adjustment pipeline for 3D reconstruction.