MDST-DGCN: A Multilevel Dynamic Spatiotemporal Directed Graph Convolutional Network for Pedestrian Trajectory Prediction

Table 3

Ablation study of MDST-DGCN-D with different level distance lists and with or without aligning operation for short-term prediction on ETH and UCY.

Level distance list	ETH	HOTEL	UNIV	ZARA1	ZARA2	AVG
{+∞}	0.870/1.792	0.490/1.011	0.626/1.273	0.407/0.867	0.333/0.703	0.545/1.129
{1, +∞}	0.862/1.772	0.465/0.989	0.532/1.139	0.402/0.862	0.324/0.687	0.517/1.090
{5, +∞}	0.853/1.757	0.453/0.931	0.624/1.269	0.400/0.852	0.322/0.684	0.530/1.100
{1, 5, +∞}	0.859/1.749	0.437/0.900	0.547/1.161	0.402/0.860	0.320/0.684	0.513/1.071
Without aligning	0.90/1.94	1.48/2.49	0.60/1.25	0.37/0.79	0.30/0.65	0.73/1.42

The level distance list {1, 5, +∞} is the default setting, and it is used for MDST-DGCN-D without aligning operation. The errors reported are ADE or FDE in meters. The values with the least error in the comparison model are bolded.