Reproduce “PPT: Pretraining with Pseudo-Labeled Trajectories for Motion Forecasting” on OpenResearch

About this paper

Yihong Xu, Yuan Yin, Éloi Zablocki, Tuan-Hung Vu, Alexandre Boulch, Matthieu Cord

PPT uses large collections of raw detector-tracker trajectories to pretrain motion-forecasting models, retaining their noise and behavioral variation instead of cleaning them into a single target trajectory. The approach substantially improves low-label learning and cross-domain transfer, including a Waymo Brier-FDE reduction from 5.358 to 0.585 with only 1% of labeled data and strong gains in end-to-end forecasting from imperfect perception inputs.

Key insights

• Noisy pseudo-trajectories can be useful supervision: exposure to localization errors, missed detections, and association failures helps models handle the imperfect inputs encountered in real systems.
• Trajectory diversity matters alongside label quality. Combining outputs from multiple detectors consistently improves performance, indicating that varied motion patterns and sensing conditions act as useful regularization.
• Much of the pretraining benefit comes from learning motion dynamics and interactions; high-definition maps and trajectory post-processing are not essential during pretraining, though maps can still help during fine-tuning.

Summary by alphaXiv.