Learning-Deep-Learning

M2I: From Factored Marginal Trajectory Prediction to Interactive Prediction

November 2022

tl;dr: Factor marginal prediction into conditional predictions, by classifying agent relationship into influencers and reactors.

Overall impression

Trajectory prediction is used by autonomous driving cars to infer future motions of nearby agents and identify risky scenarios to enable safe driving. DL based methods excel at prediction marginal trajectories for single agents, but it remains an open problem to jointly predict scene compliant trajectories over multiple agents.

Even DesnseTNT cannot handle two agents, hundred goals for each agent.

Key ideas

Technical details

Notes