Recover to Predict: Progressive Retrospective Learning for Variable-Length Trajectory Prediction
This addresses a practical challenge in autonomous driving by enabling more robust predictions with incomplete data, though it is incremental as it builds on existing methods.
The paper tackles the problem of variable-length, incomplete observations in trajectory prediction for autonomous driving by proposing a Progressive Retrospective Framework (PRF) that gradually aligns features from incomplete to complete observations, achieving improved accuracy on Argoverse datasets.
Trajectory prediction is critical for autonomous driving, enabling safe and efficient planning in dense, dynamic traffic. Most existing methods optimize prediction accuracy under fixed-length observations. However, real-world driving often yields variable-length, incomplete observations, posing a challenge to these methods. A common strategy is to directly map features from incomplete observations to those from complete ones. This one-shot mapping, however, struggles to learn accurate representations for short trajectories due to significant information gaps. To address this issue, we propose a Progressive Retrospective Framework (PRF), which gradually aligns features from incomplete observations with those from complete ones via a cascade of retrospective units. Each unit consists of a Retrospective Distillation Module (RDM) and a Retrospective Prediction Module (RPM), where RDM distills features and RPM recovers previous timesteps using the distilled features. Moreover, we propose a Rolling-Start Training Strategy (RSTS) that enhances data efficiency during PRF training. PRF is plug-and-play with existing methods. Extensive experiments on datasets Argoverse 2 and Argoverse 1 demonstrate the effectiveness of PRF. Code is available at https://github.com/zhouhao94/PRF.