MoFlow: One-Step Flow Matching for Human Trajectory Forecasting via Implicit Maximum Likelihood Estimation based Distillation
MoFlow is a new motion prediction model that predicts multiple future paths for all agents in a scene simultaneously. It uses a novel loss function to encourage diverse predictions, addressing human motion's uncertainty. By using Implicit Maximum Likelihood Estimation (IMLE) distillation, the model can generate these complex predictions in just one step, making it significantly faster than previous methods.