Reproduce “Trajectory prediction for heterogeneous agents: A performance analysis on small and imbalanced datasets” on OpenResearch

About this paper

Tiago Rodrigues de Almeida, Yufei Zhu, Andrey Rudenko, Tomasz P. Kucner, Johannes A. Stork, Martin Magnusson, Achim J. Lilienthal

This paper systematically analyzes class-conditioned trajectory prediction methods for heterogeneous agents, evaluating their performance across varying data amounts and class distributions. It finds that incorporating agent class information consistently enhances prediction accuracy, with pattern-based methods demonstrating superior robustness in limited or imbalanced data scenarios, while deep learning approaches excel with sufficient and balanced datasets.

Key insights

• Incorporating agent class information into trajectory prediction models consistently improves accuracy across both pattern-based and deep learning methodologies.
• Pattern-based methods (cMoD) exhibit superior robustness and stability when faced with limited training data or imbalanced class distributions, making them suitable for 'cold-start' robotics scenarios.
• Deep learning methods, particularly class-conditioned generative models, perform best with sufficient and balanced data, effectively leveraging larger datasets to generate multiple plausible trajectories.

Summary by alphaXiv.