The Argoverse Trajectory Retrieval Benchmark

Eric Zhan, Jagjeet Singh, Yisong Yue, Andrew Hartnett

Keywords: information retrieval, autonomous vehicles, trajectory embeddings

TL;DR: We introduce a new dataset and benchmark for contextual multi-intent trajectory retrieval of driving scenarios.

Abstract: As tracking data becomes more readily available in many domains such as sports, animal tracking, and autonomous vehicles, so does the need for effective information access and retrieval of those growing datasets. To that end, we develop the Argoverse Trajectory Retrieval Benchmark for contextual trajectory retrieval of driving scenarios. The goal of this task is to find similar trajectories from within a large dataset given a query trajectory. This task is challenging because there are many dimensions of variation in which two trajectories can be similar, such as vehicle kinematics, social causality, and road configurations. To our knowledge, this is the first standardized benchmark for trajectory retrieval of driving scenarios. We also provide an evaluation of baseline approaches based on representation learning and relevance feedback, and highlight several areas for improvement for which machine learning can play a large role in future work.

URL: https://github.com/ezhan94/argoverse-trajectory-retrieval-benchmark