Research map / Label-efficient and robust learning / Self-supervised and contrastive
PreTraM: Self-Supervised Pre-training via Connecting Trajectory and Map
Chenfeng Xu, Tian Li, Chen Tang, Lingfeng Sun, Kurt Keutzer, Masayoshi Tomizuka, Alireza Fathi, Wei Zhan
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| add PreTraM pre-training (TMCL + MCL) | AgentFormer · nuScenes val | FDE-10: 3.902 -> 3.687 | Improved |
| add PreTraM pre-training (TMCL + MCL) | Trajectron++ · nuScenes val | FDE-10: 3.221 -> 3.040 | Improved |
| add PreTraM pre-training | Trajectron++ · nuScenes | KDE NLL: 2.487 -> 2.380 | Improved |
| add PreTraM pre-training | Trajectron++ · nuScenes | Boundary violation rate (%): 23.7 -> 21.9 | Improved |
| add trajectory-map contrastive learning (TMCL) to MCL (ablation) | PreTraM-AgentFormer · nuScenes val (Map-CNN) | ADE-5: 2.442 -> 2.391 | Improved |
| add map contrastive learning (MCL) to TMCL (ablation) | PreTraM-AgentFormer · nuScenes val (Map-CNN) | ADE-5: 2.451 -> 2.391 | Improved |
| use TMCL instead of masked trajectory modeling alongside MCL (ablation) | PreTraM-AgentFormer · nuScenes val (Map-CNN) | FDE-5: 5.322 -> 5.177 | Improved |
| load pretrained trajectory encoder in addition to map encoder (ablation) | PreTraM-AgentFormer · nuScenes val (Map-CNN) | ADE-5: 2.454 -> 2.391 | Improved |
| load pretrained map encoder in addition to trajectory encoder (ablation) | PreTraM-AgentFormer · nuScenes val (Map-CNN) | ADE-5: 2.399 -> 2.391 | No clear effect |
| use dropout rather than rotation as MCL augmentation (ablation) | PreTraM-AgentFormer · nuScenes val (ResNet18) | ADE-5: 2.508 -> 2.372 | Improved |
| use dropout rather than gaussian noise as MCL augmentation (ablation) | PreTraM-AgentFormer · nuScenes val (ResNet18) | FDE-5: 5.601 -> 5.183 | Improved |
| train on 10% of trajectory data with PreTraM | AgentFormer (ResNet18) · nuScenes val, 10% trajectories | ADE-5: 3.818 -> 3.618 | Improved |
Closest papers by meaning
- STraj: Self-training for Bridging the Cross-Geography Gap in Trajectory Prediction (AAAI 2025 · similarity 0.87)
- AutoTraces: Autoregressive Trajectory Forecasting via Multimodal Large Language Models (CVPR 2026 · similarity 0.86)
- SingularTrajectory: Universal Trajectory Predictor Using Diffusion Model (CVPR 2024 · similarity 0.86)
- Improving Transferability for Cross-Domain Trajectory Prediction via Neural Stochastic Differential Equation (AAAI 2024 · similarity 0.85)
- SmartPretrain: Model-Agnostic and Dataset-Agnostic Representation Learning for Motion Prediction (ICLR 2025 · similarity 0.85)
- TrajMamba: An Efficient and Semantic-rich Vehicle Trajectory Pre-training Model (NeurIPS 2025 · similarity 0.84)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.
