Research map / Probabilistic and scientific modeling / Molecules and proteins
C2F-TP: A Coarse-to-Fine Denoising Framework for Uncertainty-Aware 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 C2F-TP, a two-stage coarse-to-fine vehicle trajectory predictor that first learns a multimodal Gaussian trajectory distribution with wave-superposition interaction pooling, then refines sampled noisy trajectories with a conditional diffusion-style denoiser.
Problem
Vehicle trajectory prediction on highways is hard because of inherent trajectory uncertainty from driving intentions and scenarios, and existing methods model inter-vehicle interactions and multimodal intentions poorly.
Approach
Combine a wave-superposition interaction pooling and re-weighted multimodal maneuver predictor that outputs a coarse trajectory distribution, then sample k noisy trajectories and refine them with a conditional denoising model conditioned on historical trajectories, reducing uncertainty via stepwise denoising.
Key results
- Stage-one output C2F-TP (C) reaches average 5s RMSE of 1.75 on NGSIM and 1.14 on highD, beating the best baseline STDAN (1.87 / 1.16).
- Full C2F-TP achieves average ADE/FDE of 0.79/1.09 m on NGSIM and 0.35/0.43 m on highD at 5s horizon, outperforming the listed baselines including LED and WSiP.
- The paper reports that C2F-TP outperforms the best baseline by 6.7%-40.4% in ADE and 13.0%-75.2% in FDE on highD.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Add the refinement module to C2F-TP (C) (ablation) | C2F-TP · NGSIM, 5s average | ADE: 0.81 -> 0.79 | Helped |
| Add the refinement module to C2F-TP (C) (ablation) | C2F-TP · NGSIM, 5s average | FDE: 1.66 -> 1.09 | Helped |
| Add the refinement module to C2F-TP (C) (ablation) | C2F-TP · HighD, 5s average | ADE: 0.43 -> 0.35 | Helped |
| Add the refinement module to C2F-TP (C) (ablation) | C2F-TP · HighD, 5s average | FDE: 1.07 -> 0.43 | Helped |
| Add interaction pooling (ablation) | C2F-TP · NGSIM and HighD | trajectory prediction performance | Helped |
| Add the re-weighted multimodal trajectory predictor (ablation) | C2F-TP · NGSIM and HighD | trajectory prediction performance | Helped |
| Add the refinement module (ablation) | C2F-TP · NGSIM and HighD | trajectory prediction performance | Helped |
| Use 10 denoising steps rather than 3 | C2F-TP · NGSIM | inference time: 186 -> 269 | Hurt |
| Use 15 denoising steps rather than 10 | C2F-TP · NGSIM | inference time: 269 -> 288 | Hurt |
| Increase denoising steps when the step count is too small | C2F-TP · NGSIM | trajectory prediction performance | Helped |
Surprising findings
- Ablation and tables suggest the refinement stage's gains on highD are large relative to the first stage, though the paper's Table 2 shows C2F-TP is not best on NGSIM FDE at all horizons.
Limitations
- Refinement stage is only evaluated on NGSIM and highD highway data, not urban scenarios.
- Inference time grows with denoising steps (186 ms to 288 ms), which may limit real-time use.
- C2F-TP is not best on NGSIM FDE at every horizon (e.g., LED is better at 4s and 5s on NGSIM FDE for some settings).
Experimental setup
Datasets: NGSIM highD
Benchmarks: NGSIM highD
Metrics: RMSE ADE FDE
Baselines: V-LSTM S-LSTM CS-LSTM S-GAN PiP-noPlan STDAN WSiP LED
Builds on: Wave-MLP WSiP DDPM DDIM LED
Closest papers by meaning
- Certified Human Trajectory Prediction (CVPR 2025 · similarity 0.88)
- Unified Uncertainty-Aware Diffusion for Multi-Agent Trajectory Modeling (CVPR 2025 · similarity 0.86)
- CCTR: Calibrating Trajectory Prediction for Uncertainty-Aware Motion Planning in Autonomous Driving (AAAI 2024 · similarity 0.86)
- CaDeT: a Causal Disentanglement Approach for Robust Trajectory Prediction in Autonomous Driving (CVPR 2024 · similarity 0.85)
- Improving Transferability for Cross-Domain Trajectory Prediction via Neural Stochastic Differential Equation (AAAI 2024 · similarity 0.85)
- LaKD: Length-agnostic Knowledge Distillation for Trajectory Prediction with Any Length Observations (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.