GTC Silicon Valley-2019 ID:S9623:GPU-Accelerated 3D Point Cloud Processing with Hierarchical Gaussian Mixtures
We'll discuss work by NVIDIA Research on processing 3D point clouds with a new type of statistical representation, the Hierarchical Gaussian Mixture. This generative model provides a compact and scalable statistical representation for 3D point cloud data that is amenable to GPU parallelization, allowing real-time computation over tens of thousands of points on low-power Tegra hardware. Our talk will cover how to efficiently create, sample from, and use these models to solve problems pertinent to autonomous vehicles and robotics. Specifically, we will outline their use in recent state-of-the-art solutions to point cloud registration and occlusion estimation.