Reproduce “Towards Generalizable Trajectory Prediction Using Dual-Level Representation Learning And Adaptive Prompting” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Kaouther Messaoud, Matthieu Cord, Alexandre Alahi
PerReg+ is a new framework that uses dual-level representation learning and adaptive prompt tuning to enhance generalizability, accuracy, and efficiency in vehicle trajectory prediction. It sets new state-of-the-art performance on nuScenes, Argoverse 2, and Waymo Open Motion Dataset benchmarks, reducing B-FDE by up to 13% with pre-training.
Key insights
- Fine-grained masking combined with self-distillation effectively forces the model to learn robust, detailed scene representations from subtle contextual cues.
- Register queries provide a structured memory for intermediate scene representations, significantly improving multimodal prediction and retaining learned dynamics.
- Adaptive prompt tuning, when guided by scene clustering, enables efficient and targeted adaptation to new scenarios while preserving generalizable features from pre-trained models.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.