3DMatch: Learning Local Geometric Descriptors from RGB-D Reconstructions This paper presents 3DMatch, a data-driven model that learns a local volumetric patch descriptor for establishing correspondences between partial 3D data. The model is trained using self-supervised feature learning from millions of correspondence labels found in existing RGB-D reconstructions. 3DMatch is able to match local geometry in new scenes for reconstruction and generalizes to different tasks and spatial scales, such as instance-level object model alignment for the Amazon Picking Challenge and mesh surface correspondence. Experiments show that 3DMatch consistently outperforms other state-of-the-art approaches by a significant margin. Code, data, benchmarks, and pre-trained models are available online at http://3dmatch.cs.princeton.edu. The paper discusses the challenges of matching local geometric features on real-world depth images due to the noisy, low-resolution, and incomplete nature of 3D scan data. Current state-of-the-art methods are typically based on histograms over geometric properties, but they are often unstable or inconsistent in real-world partial surfaces. 3DMatch is a 3D convolutional neural network (ConvNet) that takes in the local volumetric region around an arbitrary interest point on a 3D surface and computes a feature descriptor for that point. The model is trained using data from existing RGB-D reconstructions, which provide a large amount of training correspondences. The model is able to match local geometry in new scenes for reconstruction and generalizes to different tasks and spatial scales. The paper also discusses related work in computer vision and graphics, including hand-crafted 3D local descriptors, learned 2D local descriptors, learned 3D global descriptors, and learned 3D local descriptors. The paper highlights the advantages of using 3D TDF voxel grids for representing 3D data, which allows reasoning over real-world spatial scale and occluded regions. The paper also discusses the implementation details of the 3DMatch network, including the network architecture, training process, and run-time information. The paper evaluates the performance of 3DMatch on various tasks, including keypoint matching, geometric registration, and 6D object pose estimation. The results show that 3DMatch outperforms other state-of-the-art methods in these tasks. The paper also discusses the ability of 3DMatch to generalize to new domains, such as 6D object pose estimation by model alignment and correspondence labeling for 3D meshes. The paper concludes that 3DMatch is a powerful local geometric descriptor that can be used for a variety of applications.3DMatch: Learning Local Geometric Descriptors from RGB-D Reconstructions This paper presents 3DMatch, a data-driven model that learns a local volumetric patch descriptor for establishing correspondences between partial 3D data. The model is trained using self-supervised feature learning from millions of correspondence labels found in existing RGB-D reconstructions. 3DMatch is able to match local geometry in new scenes for reconstruction and generalizes to different tasks and spatial scales, such as instance-level object model alignment for the Amazon Picking Challenge and mesh surface correspondence. Experiments show that 3DMatch consistently outperforms other state-of-the-art approaches by a significant margin. Code, data, benchmarks, and pre-trained models are available online at http://3dmatch.cs.princeton.edu. The paper discusses the challenges of matching local geometric features on real-world depth images due to the noisy, low-resolution, and incomplete nature of 3D scan data. Current state-of-the-art methods are typically based on histograms over geometric properties, but they are often unstable or inconsistent in real-world partial surfaces. 3DMatch is a 3D convolutional neural network (ConvNet) that takes in the local volumetric region around an arbitrary interest point on a 3D surface and computes a feature descriptor for that point. The model is trained using data from existing RGB-D reconstructions, which provide a large amount of training correspondences. The model is able to match local geometry in new scenes for reconstruction and generalizes to different tasks and spatial scales. The paper also discusses related work in computer vision and graphics, including hand-crafted 3D local descriptors, learned 2D local descriptors, learned 3D global descriptors, and learned 3D local descriptors. The paper highlights the advantages of using 3D TDF voxel grids for representing 3D data, which allows reasoning over real-world spatial scale and occluded regions. The paper also discusses the implementation details of the 3DMatch network, including the network architecture, training process, and run-time information. The paper evaluates the performance of 3DMatch on various tasks, including keypoint matching, geometric registration, and 6D object pose estimation. The results show that 3DMatch outperforms other state-of-the-art methods in these tasks. The paper also discusses the ability of 3DMatch to generalize to new domains, such as 6D object pose estimation by model alignment and correspondence labeling for 3D meshes. The paper concludes that 3DMatch is a powerful local geometric descriptor that can be used for a variety of applications.