A2X: An Agent and Environment Interaction Benchmark for Multimodal Human Trajectory Prediction

Samuel S. Sohn, Mihee Lee, Seonghyeon Moon, Gang Qiao, Usman Muhammad, Sejong Yoon, vladimir pavlovic, Mubbasir Kapadia

Keywords: human trajectory prediction, datasets, evaluation metrics

Abstract: Recent trends in human trajectory prediction are the development of generative models which generate distributions of trajectories. However existing metrics are suited only for single (unimodal) trajectory instances. Furthermore, existing datasets are largely limited to small-scale interactions between people, with little to no agent-to-agent environment interaction. To address these challenges, we propose a dataset that compensates for the lack of agent-to-environment interaction in existing datasets with a new simulated dataset and metrics to convey model performance with more reliability and nuance. A subset of these metrics are novel multiverse metrics, which are better-suited for multimodal models than existing metrics but are still applicable to unimodal models. Our results showcase the benefits of the augmented dataset and metrics. The dataset is available at: https://mubbasir.github.io/HTP-benchmark/.

URL: https://mubbasir.github.io/HTP-benchmark/