In this work, we present a single-stage framework, named S2F2, for forecasting multiple human trajectories from raw video images by predicting future optical flows. S2F2 differs from the previous ...