PointFlow 3D Point Cloud Generation with Continuous Normalizing Flows

point cloud as a sample

• to gen. a p.c. -> gen. a dist. of dist.
• model the dist. of points as transformation of 3D points from a prior distribution
  • VI framework
  • sample from prior then move them to posterior dist.
• continuous normalizing flows
  • well…Jacobians.
• a learnable prior
• loss
  • shape encoding -> prior
  • prob. of reconstruction: MC
  • entropy of posterior