Research map / Multi-agent systems and games / Human-AI and goal-conditioned
SmartRefine: A Scenario-Adaptive Refinement Framework for Efficient Motion 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 SmartRefine, a scenario-adaptive, plug-in trajectory refinement module for motion prediction that adapts anchors, retrieval radius, and context encoding per scenario and uses a learned quality score to pick the number of refinement iterations.
Problem
Two-stage motion prediction refinement for autonomous driving is either computationally expensive (using all context) or gives limited gains, and uses fixed refinement configurations and iteration counts across all scenarios.
Approach
A decoupled refinement module that selects anchors along predicted trajectory segments, retrieves context with an iteration- and speed-dependent radius, encodes context in anchor-centric coordinates, refines segment-by-segment recurrently, and uses a predicted quality score (normalized error-based target) to adaptively stop iterations.
Key results
- Adding SmartRefine to QCNet on Argoverse 2 val reduces minFDE from 1.253 to 1.240 and MR from 0.157 to 0.156, with +0.3G Flops.
- On Argoverse 2 test, QCNet w/ Ours reaches minFDE 1.23 and is reported to outperform all published ensemble-free single-agent entries on the leaderboard at submission.
- On Argoverse val, HiVT minFDE drops from 0.969 to 0.911 when SmartRefine is added, with 0.1G extra Flops and similar latency.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| add SmartRefine | HiVT, Prophnet, and mmTransformer · Argoverse val | minFDE: 6.0%, 3.7%, 5.4%, 4.3%, 3.5% | Helped |
| add SmartRefine | DenseTNT and QCNet (no ref) · Argoverse 2 test | minFDE: 4.2%, 3.9% | Helped |
| increase anchor count from 1 to 2 (ablation) | SmartRefine on HiVT · Argoverse val | minFDE: 0.928 -> 0.911 | Helped |
| increase anchor count beyond 2 (ablation) | SmartRefine on HiVT · Argoverse val | minFDE: 0.911; 0.915; 0.916 | Mixed |
| use anchor-centric instead of agent-centric context encoding (ablation) | SmartRefine on HiVT · Argoverse val | minFDE: 0.941 -> 0.911 | Helped |
| replace fixed retrieval radius with adaptive radius (ablation) | SmartRefine on HiVT · Argoverse val | minFDE: 0.926 -> 0.911; 0.923 -> 0.911; 0.921 -> 0.911; 0.930 -> 0.911 | Helped |
| use exponential rather than linear adaptive-radius decay (ablation) | SmartRefine on HiVT · Argoverse val | minFDE: 0.911; 0.911 | No clear effect |
| use exponential rather than linear adaptive-radius decay (ablation) | SmartRefine on HiVT · Argoverse val | Flops: 245; 130 | Helped |
| use adaptive instead of fixed refinement iterations (ablation) | SmartRefine on HiVT · Argoverse and Argoverse 2 | minFDE | Helped |
| continue refining already high-quality trajectories (ablation) | SmartRefine on HiVT · Argoverse val | prediction quality score | Hurt |
Surprising findings
- Not every trajectory benefits from refinement: initially high-quality trajectories can be pushed away from ground truth by over-refinement, while low-quality ones improve.
Limitations
- Gains on QCNet are smaller because it already includes a large refinement network
- Only single-agent prediction is considered; multi-agent joint prediction is future work
- Quality score threshold and max iteration are hyperparameters requiring tuning
Experimental setup
Datasets: Argoverse 1 Argoverse 2
Benchmarks: Argoverse 1 Argoverse 2
Metrics: minADE minFDE MR Flops latency parameters
Baselines: HiVT Prophnet mmTransformer DenseTNT QCNet QCNet (no ref) MTR R-Pred
Builds on: QCNet MTR Multipath GRU
Closest papers by meaning
- SMART: Scalable Multi-agent Real-time Motion Generation via Next-token Prediction (NeurIPS 2024 · similarity 0.88)
- SmartPretrain: Model-Agnostic and Dataset-Agnostic Representation Learning for Motion Prediction (ICLR 2025 · similarity 0.88)
- DeMo: Decoupling Motion Forecasting into Directional Intentions and Dynamic States (NeurIPS 2024 · similarity 0.87)
- Motion Forecasting in Continuous Driving (NeurIPS 2024 · similarity 0.87)
- HPNet: Dynamic Trajectory Forecasting with Historical Prediction Attention (CVPR 2024 · similarity 0.86)
- UniMotion: A Unified Motion Framework for Simulation, Prediction and Planning (NeurIPS 2025 · similarity 0.86)
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.