3DMatch: Learning Local Geometric Descriptors from RGB-D Reconstructions

Abstract:
Matching local geometric features on real-world depth images is a challenging task due to the noisy, low-resolution, and incomplete nature of 3D scan 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 present 3DMatch, a data-driven model that learns a local volumetric patch descriptor for establishing correspondences between partial 3D data. To amass training data for our model, we propose an unsupervised feature learning method that leverages the millions of correspondence labels found in existing RGB-D reconstructions. Experiments show that our descriptor is not only able to match local geometry in new scenes for reconstruction, but also generalize to different tasks and spatial scales (e.g. instance-level object model alignment for the Amazon Picking Challenge, and mesh surface correspondence). Results show that 3DMatch consistently outperforms other state-of-the-art approaches by a significant margin.

[theincubus]

"Given an interest point on a 3D surface..." at t=0:57 is a bit hand-wavy. This research is excellent (and long awaited by researchers like me who only dabble in 3D vision), so I will be reading the paper, but can you elaborate? How are these points chosen? Also, you say you train a Siamese style network. For what purpose? Is one channel depth such as HHA/DHA and the other color?