PerReg+: Towards Generalizable Trajectory Prediction using Dual-Level Representation Learning and Adaptive Prompting

Official codebase for the CVPR 2025 paper:
"Towards Generalizable Trajectory Prediction using Dual-Level Representation Learning and Adaptive Prompting"
by Kaouther Messaoud, Matthieu Cord, and Alexandre Alahi.

---

Table of Contents

• Overview
• Paper
• Highlights
• Benchmarks
• Coming Soon

---

Overview

PerReg+ is a novel transformer-based framework for vehicle trajectory prediction. It addresses key challenges in autonomous driving by:

• Generalizing across domains and datasets
• Enabling multimodal prediction without clustering or NMS
• Providing efficient domain adaptation via adaptive prompt tuning

Key Innovations:

• Dual-Level Representation Learning using Self-Distillation (SD) and Masked Reconstruction (MR)
• Register-based Queries for efficient and structured multimodal output
• Segment-Level Reconstruction of trajectories and lanes
• Adaptive Prompt Tuning for scalable and fast fine-tuning

PerReg+ achieves state-of-the-art results on the UniTraj Benchmark and shows strong cross-domain generalization.

---

Paper

• 📄 CVPR 2025 Paper (PDF) (link will be added when available)

---

Highlights

• 📉 6.8% reduction in Brier-minFDE on small datasets through SSL pretraining
• 🔄 11.8% improvement in cross-domain generalization performance
• ⚡ Prompt-based fine-tuning: adapts with minimal overhead
• 🛠 No clustering/NMS required for multimodal prediction
• 🧠 Richer scene understanding through segment-level reconstruction and dual-level supervision

---

Benchmarks

Evaluated on:

• nuScenes
• Argoverse 2
• Waymo Open Motion Dataset (WOMD)

Achievements:

• Outperforms AutoBot, MTR, and Forecast-MAE baselines
• Best B-FDE and minFDE across single-dataset and multi-dataset training
• Strong generalization to unseen domains

---

Coming Soon

• 📦 Full code release (training & evaluation)
• 🏆 Pretrained model checkpoints
• 📚 Detailed instructions and tutorials