In-Depth Comparison of Regularization Methods For Long-Tailed Learning in Trajectory Prediction

TL;DR — We compare regularization-based long-tailed learning techniques for trajectory prediction, and provide in-depth analysis.

Abstract

Autonomous robots have the biggest potential for risk because they operate in open-ended environments where humans interact in complex, diverse ways. To operate, such systems must predict this behaviour, especially if it's part of the unexpected and potentially dangerous long tail of the dataset. Previous works on long-tailed trajectory prediction use models which do not predict a distribution of trajectories with likelihoods associated with each prediction. Furthermore, they report metrics which are biased by the ground-truth. Therefore, we aim to examine regularization methods for long-tailed trajectory prediction by comparing them on the KDE metric, which is designed to compare distributions of trajectories. Moreover, we are the first to report the performance of these methods on both the pedestrian and vehicle classes of the NuScenes dataset.