Research map / Label-efficient and robust learning / Recognition and long-tail
Improving Transferability for Cross-Domain Trajectory Prediction via Neural Stochastic Differential Equation
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 TrajSDE, a neural stochastic differential equation encoder-decoder with dataset-specific diffusion networks and SDE-Net-style uncertainty training, to transfer multi-agent trajectory prediction across datasets with different time configurations and tracklet errors.
Problem
Trajectory prediction models trained on large-scale datasets transfer poorly to small datasets because of discrepancies in data acquisition, namely differing history/prediction lengths, sampling rates, and detector/tracker errors.
Approach
Replace discrete temporal encoders/decoders with an ODE-RNN-style SDE model that handles arbitrary time steps continuously, share a drift network across datasets while using separate dataset-specific diffusion networks trained with SDE-Net in-distribution/out-of-distribution objectives so tracklet-noisy inputs receive larger Brownian weighting.
Key results
- On nuScenes val with N+A training, HiVT + ours reaches mADE10 0.913 vs 1.045 for nuScenes-only (12.55% gain) and 0.966 for vanilla HiVT with N+A.
- For goal-conditioned MUSE-VAE, the gain from adding INTERACTION on Lyft val rises from 8.90% (vanilla) to 30.56% with the proposed method.
- With no target training data, the baseline HiVT diverges above 14 mADE10 while the proposed method stays at 1.26.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| train HiVT with nuScenes and Argoverse using the proposed method instead of vanilla HiVT | HiVT · nuScenes val | mADE10: 0.966 -> 0.913 | Helped |
| train HiVT with nuScenes and WOMD using the proposed method instead of vanilla HiVT | HiVT · nuScenes val | mADE10: 0.950 -> 0.893 | Helped |
| apply the proposed framework when training MUSE-VAE with nuScenes and INTERACTION | MUSE-VAE · nuScenes val | mADE10: 2.178 -> 1.953 | Helped |
| apply the proposed framework when training MUSE-VAE with Lyft and INTERACTION | MUSE-VAE · Lyft val | mADE10: 1.073 -> 0.827 | Helped |
| apply the proposed method when training HiVT with 5% WOMD and Argoverse | HiVT · WOMD val | mADE10: 0.9286 -> 0.8496 | Helped |
| add the proposed method to HiVT for uncertain tracklet samples | HiVT · nuScenes val, ood samples | mADE10: 1.414 -> 1.333 | Helped |
| use a sampled Gaussian latent for the past feature (ablation) | HiVT + ours · nuScenes val, train N+A | mADE10: 0.913 -> 1.225 | Hurt |
| keep encoder drift and diffusion networks at four layers rather than reducing them to two (ablation) | HiVT + ours · nuScenes val, train N+A | mADE10: 0.927 -> 0.913 | Helped |
| keep decoder drift and diffusion networks at four layers rather than reducing them to two (ablation) | HiVT + ours · nuScenes val, train N+A | mADE10: 0.918 -> 0.913 | Helped |
| share the drift network across datasets (ablation) | HiVT + ours · nuScenes val, train N+A | mADE10: 0.934 -> 0.913 | Helped |
| use separate diffusion networks across datasets (ablation) | HiVT + ours · nuScenes val, train N+A | mADE10: 0.944 -> 0.913 | Helped |
| include the uncertainty training objective (ablation) | HiVT + ours · nuScenes val, train N+A | mADE10: 0.940 -> 0.913 | Helped |
Surprising findings
- Gaussian latent modeling of the past feature severely worsens performance (1.225 vs 0.913 mADE10).
- Plain latentSDE gives a smaller gain than ODE-RNN, attributed to a single diffusion network failing to handle heterogeneous tracklet errors.
Limitations
- Only vehicle trajectories are trained and evaluated.
- Tracklet noise is simulated with Gaussian perturbation of ego-agent trajectories rather than real detector/tracker errors.
Experimental setup
Datasets: nuScenes Argoverse Waymo Open Motion Dataset Lyft INTERACTION
Benchmarks: nuScenes Argoverse WOMD Lyft INTERACTION
Metrics: mADE10
Baselines: HiVT MUSE-VAE ODE-RNN latentSDE random dropping feature alignment domain adaptation
Builds on: Neural ODE ODE-RNN SDE-Net latentSDE HiVT MUSE-VAE
Closest papers by meaning
- STraj: Self-training for Bridging the Cross-Geography Gap in Trajectory Prediction (AAAI 2025 · similarity 0.87)
- Unified Uncertainty-Aware Diffusion for Multi-Agent Trajectory Modeling (CVPR 2025 · similarity 0.86)
- HPNet: Dynamic Trajectory Forecasting with Historical Prediction Attention (CVPR 2024 · similarity 0.86)
- SingularTrajectory: Universal Trajectory Predictor Using Diffusion Model (CVPR 2024 · similarity 0.86)
- Certified Human Trajectory Prediction (CVPR 2025 · similarity 0.85)
- C2F-TP: A Coarse-to-Fine Denoising Framework for Uncertainty-Aware Trajectory Prediction (AAAI 2025 · 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.