Understanding the Task and Data Misconceptions in Online Map Based Motion Prediction for Autonomous Driving and a Boundary-Free Baseline

Zhiyuan Zhang, Xiaosong Jia, Junchi Yan

Keywords: Motion Prediction

Abstract:

In autonomous driving (AD), online high-definition (HD) map estimation is gaining increasing attention. To examine how online-estimated HD maps impact downstream tasks, the protocol of online mapping based motion prediction emerges. This protocol follows a two-stage training paradigm: online mapping models are firstly trained and then used to output map elements which are fed as inputs for motion prediction models. In this paper, we conduct in-depth study to investigate the challenges and misunderstandings associated with the protocol and propose OMMP-Bench, a well-defined and insightful benchmark of online map based motion prediction. We identify that the current dataset splits are unsuitable for two-stage training, leading to a severe train-validation gap, and thus we design a new data partitioning split. Furthermore, we find that the perception range of map prediction models does not fully meet the requirements of motion prediction, resulting in a lack of map elements for agents far from the ego vehicle. This issue is obscured by incorrect metrics that evaluate only the ego vehicle’s trajectory. We address it by refining the metrics to evaluate all moving vehicles and separately report performance for agents under different distance ranges. Further, to alleviate the issue of missing map elements for faraway agents, we introduce a new baseline that directly uses image features generated by the online mapping model. These features are not constrained by perception range and could supplement environmental information around agents beyond the online map’s coverage. We further explore how different map elements influence motion prediction, as existing online mapping models have different designs of output format.

We conduct thorough experiments to verify the proposed corrections. We will open source the related code and checkpoints. We hope OMMP-Bench could solve the long-standing mis-usage and misunderstanding of the emerging field and provide insights for further co-development of online mapping and motion prediction models.

Primary Area: datasets and benchmarks

Submission Number: 6901