We propose APML, a differentiable and efficient loss for point cloud prediction tasks, approximating one-to-one matching using Sinkhorn iterations with adaptive temperature.

We provide the statistical guarantee for Sinkhorn bridge method to estimate Schrödinger bridge