Research map / Probabilistic and scientific modeling / Uncertainty and Gaussian processes
Producing and Leveraging Online Map Uncertainty 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.
Extends MapTR, MapTRv2, and StreamMapNet to output Laplace-distributed vertex uncertainty and feeds it into DenseTNT and HiVT trajectory predictors, reporting faster convergence and better prediction on nuScenes.
Problem
Online HD map estimation methods output no uncertainty or confidence, so downstream trajectory prediction implicitly treats map elements as certain and fails when mapping errors occur.
Approach
A general vectorized map uncertainty formulation that replaces the regression head with per-vertex Laplace distributions trained by NLL loss, and encodes the resulting location, scale, and class probabilities as inputs to map-encoder trajectory predictors.
Key results
- Incorporating map uncertainty reduces HiVT miss rate with MapTR by 15% (0.0981 to 0.0834) and improves DenseTNT and HiVT across most of 16 mapping/prediction combinations.
- DenseTNT training converges in 4 epochs instead of 8 with MapTR map uncertainty (50% faster).
- DenseTNT with MapTRv2-Centerline plus uncertainty achieves minADE 0.8135 versus 0.8809 with GT map.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Add map uncertainty to MapTR map inputs (ablation) | HiVT · nuScenes | minADE: 0.4015 -> 0.3854 | Helped |
| Add map uncertainty to MapTR map inputs (ablation) | HiVT · nuScenes | minFDE: 0.8418 -> 0.7909 | Helped |
| Add map uncertainty to MapTR map inputs (ablation) | HiVT · nuScenes | MR: 0.0981 -> 0.0834 | Helped |
| Add map uncertainty to MapTRv2 map inputs (ablation) | HiVT · nuScenes | minADE: 0.4057 -> 0.3930 | Helped |
| Add map uncertainty to MapTRv2 map inputs (ablation) | HiVT · nuScenes | minFDE: 0.8499 -> 0.8127 | Helped |
| Add map uncertainty to MapTRv2 map inputs (ablation) | HiVT · nuScenes | MR: 0.0992 -> 0.0857 | Helped |
| Add map uncertainty to MapTRv2-Centerline map inputs (ablation) | HiVT · nuScenes | minADE: 0.3790 -> 0.3727 | Helped |
| Add map uncertainty to MapTRv2-Centerline map inputs (ablation) | HiVT · nuScenes | MR: 0.0853 -> 0.0726 | Helped |
| Add map uncertainty to MapTRv2-Centerline map inputs (ablation) | DenseTNT · nuScenes | MR: 0.1520 -> 0.1593 | Hurt |
| Add map uncertainty to StreamMapNet map inputs (ablation) | DenseTNT · nuScenes | minADE: 0.9492 -> 0.9036 | Helped |
| Add map uncertainty to StreamMapNet map inputs (ablation) | DenseTNT · nuScenes | MR: 0.2569 -> 0.2359 | Helped |
| Add map uncertainty to DenseTNT inputs (ablation) | DenseTNT + MapTR · nuScenes | epochs to convergence: 8 -> 4 (−50%) | Helped |
Surprising findings
- Accuracy improvements in upstream map estimation (MapTRv2 over MapTR) did not translate into better downstream prediction accuracy.
- DenseTNT trained on MapTRv2-Centerline maps outperforms DenseTNT trained on ground-truth lane centerlines, attributed to multiple centerlines per lane giving richer endpoint options.
Limitations
- Experiments limited to nuScenes, with little high-speed driving data.
- Uncertainty calibration is not evaluated; leveraging it for map calibration is left to future work due to fuzzy point set matching.
- Only regression and classification heads are modified; other uncertainty sources such as connectivity are not modeled.
Experimental setup
Datasets: nuScenes
Benchmarks: nuScenes
Metrics: mAP (Chamfer distance thresholds 0.5/1.0/1.5 m) minADE minFDE Miss Rate epochs to convergence
Baselines: MapTR MapTRv2 MapTRv2-Centerline StreamMapNet GT Map
Builds on: MapTR MapTRv2 StreamMapNet DenseTNT HiVT
Closest papers by meaning
- Leveraging SD Map to Augment HD Map-based Trajectory Prediction (CVPR 2025 · similarity 0.89)
- SDTagNet: Leveraging Text-Annotated Navigation Maps for Online HD Map Construction (NeurIPS 2025 · similarity 0.86)
- MapExpert: Online HD Map Construction with Simple and Efficient Sparse Map Element Expert (AAAI 2025 · similarity 0.86)
- Beyond Chamfer Distance: Granular Order-aware Evaluation Metric For Online Mapping (NeurIPS 2026 · similarity 0.86)
- C2F-TP: A Coarse-to-Fine Denoising Framework for Uncertainty-Aware Trajectory Prediction (AAAI 2025 · similarity 0.85)
- Future-Aware End-to-End Driving: Bidirectional Modeling of Trajectory Planning and Scene Evolution (NeurIPS 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.