Pure geometric 3D vision has advanced to become fast enough and robust enough to power SLAM and visual odometry for AR and VR systems. Meanwhile, machine learning, in particular deep learning, has advanced to achieve human-level performance on many complex vision tasks such as recognition and segmentation. But for the most part, these areas have remained separate. What lies at their intersection? With the advent of increasing on-device compute, specialized ASICs for machine learning, and more sophisticated machine learning approaches themselves, there is now an opportunity to take existing geometric vision systems and upgrade them with ML. In this talk, we’ll explore a few of these existing applications, and how they will evolve once they are combined with a machine learning approach that is built specifically to consume and analyze 3D data.