3DMatch is a deep learning-based local volumetric patch descriptor that operates on 3D data including point clouds, depth maps, and meshes. It uses convolutional neural networks to learn geometric features for establishing correspondences between partial 3D scans. The toolbox provides both inference code for geometric registration and training code to learn the descriptor from existing RGB-D reconstructions. It is primarily used for 3D reconstruction tasks involving noisy and incomplete scan data.