Supplementary Material

Section A provides additional details for the method. The scene encoder is to extract the environment information. Following [19], we sample the frames at 2.5 HZ and predict future For ETH and UCY datasets, we adopt the standard metrics ( i . Due to the limitations discussed in Section 4.1, we introudce curve smoothing (CS) into current We conduct experiments on P A V using the traditional ADE/FDE metrics. In particular, our method improves the FDE by 13.6% on PETS.