Reproduce “EigenTrajectory: Low-Rank Descriptors for Multi-Modal Trajectory Forecasting” on OpenResearch

About this paper

Inhwan Bae, Jean Oh, Hae-Gon Jeon

EigenTrajectory (ET) is a data-driven, low-rank descriptor for human trajectories that learns principal motion components using Singular Value Decomposition. This approach consistently improves trajectory prediction accuracy, reducing Final Displacement Error by up to 74.3% for PECNet and 49.2% on average for Social-STGCNN, and enhances reliability across multiple state-of-the-art models by enabling all forecasting operations within a compact, learned feature space.

Key insights

• A data-driven, low-rank basis learned via SVD provides a more suitable and compact abstraction for human trajectory dynamics than predefined parametric functions.
• Operating the entire trajectory forecasting pipeline within this lower-dimensional, data-derived 'ET space' significantly improves prediction accuracy, efficiency, and robustness to noise.
• Clustering normalized trajectories in the ET space efficiently generates diverse and reliable 'trajectory anchors' for multi-modal forecasting, outperforming clustering in high-dimensional Euclidean space.

Summary by alphaXiv.