Research map / Label-efficient and robust learning / Recognition and long-tail
STraj: Self-training for Bridging the Cross-Geography Gap in Trajectory Prediction
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes STraj, a plug-and-play self-training UDA pipeline (complementary agent/map masking pre-training, consistency-based pseudo-label update, trajectory-induced contrastive learning) to close the cross-geography gap in trajectory prediction on Argoverse 1.
Problem
Trajectory predictors trained in one city (geography) degrade sharply when deployed in another because agent and lane-segment distributions differ; annotating target-city trajectories is costly.
Approach
Adapts pseudo-label self-training to trajectory prediction by (1) pre-training with distance-weighted complementary agent and map masking, (2) updating pseudo trajectories by cosine-similarity consistency with historical epochs and weighting the loss by consistency, and (3) using consistency-selected positives/negatives across domains for contrastive learning on agent representations.
Key results
- On MIA->PIT with LaneGCN (K=1), STraj reduces MR from 0.755 (Source Only) to 0.685 and improves minADE/minFDE by about 15.5% and 18.5%.
- Ablation on MIA->PIT LaneGCN: augmentation, pseudo label update, and trajectory-induced contrastive learning together bring minADE from 2.32 to 1.96 and minFDE from 5.46 to 4.45.
- Cosine similarity of whole trajectories for consistency outperforms Euclidean-distance variants in ablation (minADE 1.96 vs 2.02 for best variant).
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| add STraj self-training | LaneGCN · MIA → PIT | minADE, K=1: 2.32 -> 1.96 | Helped |
| add STraj self-training | LaneGCN · PIT → MIA | minADE, K=1: 2.68 -> 2.89 | Hurt |
| add complementary agent and map augmentations (ablation) | LaneGCN · MIA → PIT | minADE, K=1: 2.32 -> 2.26 | Helped |
| add pseudo label update strategy (ablation) | LaneGCN with augmentation · MIA → PIT | minADE, K=1: 2.26 -> 2.09 | Helped |
| add trajectory-induced contrastive learning (ablation) | LaneGCN with augmentation and PLUS · MIA → PIT | minADE, K=1: 2.09 -> 1.96 | Helped |
| use agent augmentation only instead of complementary augmentation (ablation) | LaneGCN · MIA → PIT | minADE, K=1: 1.96 -> 2.11 | Hurt |
| use map augmentation only instead of complementary augmentation (ablation) | LaneGCN · MIA → PIT | minADE, K=1: 1.96 -> 2.05 | Hurt |
| replace weighted random masking with random masking (ablation) | LaneGCN augmentation · MIA → PIT | minADE, K=1: 1.96 -> 2.19 | Hurt |
| measure trajectory consistency with Euclidean distance rather than cosine similarity (ablation) | LaneGCN pseudo label update · MIA → PIT | minADE, K=1: 1.96 -> 2.02 | Hurt |
| combine Euclidean distance and cosine similarity for trajectory consistency (ablation) | LaneGCN pseudo label update · MIA → PIT | minADE, K=1: 1.96 -> 2.20 | Hurt |
| calculate cosine similarity only at trajectory endpoints (ablation) | LaneGCN pseudo label update · MIA → PIT | minADE, K=1: 1.96 -> 2.21 | Hurt |
| calculate Euclidean distance only at trajectory endpoints (ablation) | LaneGCN pseudo label update · MIA → PIT | minADE, K=1: 1.96 -> 2.16 | Hurt |
Surprising findings
- Off-the-shelf image self-training method ICON degrades trajectory prediction performance in some cross-city settings (e.g., MIA->PIT LaneGCN K=1 minADE 2.51 vs Source Only 2.32).
Limitations
- Evaluated only on Argoverse 1 with two cities, so generality across more geographies is unshown.
- Relies on hand-set thresholds (consistency and confidence) and scaling hyperparameters.
Experimental setup
Datasets: Argoverse 1
Benchmarks: Argoverse 1 (MIA->PIT, PIT->MIA)
Metrics: minADE minFDE MR
Baselines: Source Only Oracle ICON S-attack Frenét
Builds on: LaneGCN ICON InfoNCE teacher forcing
Closest papers by meaning
- Enhancing Trajectory Prediction through Self-Supervised Waypoint Distortion Prediction (ICML 2024 · similarity 0.87)
- Improving Transferability for Cross-Domain Trajectory Prediction via Neural Stochastic Differential Equation (AAAI 2024 · similarity 0.87)
- HPNet: Dynamic Trajectory Forecasting with Historical Prediction Attention (CVPR 2024 · similarity 0.87)
- UniTraj: Learning a Universal Trajectory Foundation Model from Billion-Scale Worldwide Traces (NeurIPS 2025 · similarity 0.86)
- Trajectory Mamba: Efficient Attention-Mamba Forecasting Model Based on Selective SSM (CVPR 2025 · similarity 0.86)
- DeMo: Decoupling Motion Forecasting into Directional Intentions and Dynamic States (NeurIPS 2024 · similarity 0.85)
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.