Research map / Probabilistic and scientific modeling / Uncertainty and Gaussian processes
Quantifying Uncertainty in Motion Prediction with Variational Bayesian Mixture
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 SeNeVA, a variational Bayesian mixture generative model for single-agent trajectory prediction that quantifies uncertainty, detects OOD inputs via entropy, and matches state-of-the-art accuracy on INTERACTION and Argoverse 2.
Problem
Motion prediction for autonomous vehicles needs multi-modal future trajectory forecasts together with well-calibrated uncertainty estimates, which most existing sequential and goal-based models do not provide.
Approach
Model the future trajectory distribution as a mixture of K disjoint Gaussian components with latent temporal variables trained by variational inference (ELBO), plus a separate assignment network that predicts mixture weights from the scene context alone, and an NMS-based destination sampler to produce a small set of representative trajectories.
Key results
- On the INTERACTION test set, SeNeVA reaches 0.446 minFDE6, 0.203 minADE6 and 5.3% miss rate with 1.3M parameters.
- On the INTERACTION val set, SeNeVA achieves 0.431 minFDE6 and 0.197 minADE6, a 35.7% minFDE6 improvement over the second-best method.
- Predictive entropy is higher for OOD cases in merging (+1.5%) and roundabout (+0.6%) scenarios, with inference at 51ms versus 1,590ms for Deep Ensembles.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| SeNeVA versus HDGT | INTERACTION test · INTERACTION test | minFDE6: 0.478 -> 0.446 | Helped |
| SeNeVA versus Multi-Branch SS-ASP | INTERACTION test · INTERACTION test | minADE6: 0.178 -> 0.203 | Hurt |
| SeNeVA versus MultiModalTransformer | INTERACTION test · INTERACTION test | MR: 0.051 -> 0.053 | Hurt |
| SeNeVA versus TNT | INTERACTION val · INTERACTION val | minFDE6: 0.670 -> 0.431 | Helped |
| SeNeVA versus TNT | INTERACTION val · INTERACTION val | minADE6: 0.210 -> 0.197 | Helped |
| SeNeVA versus Forecast-MAE | Argoverse 2 val · Argoverse 2 val | minFDE6: 1.409 -> 1.319 | Helped |
| SeNeVA versus Forecast-MAE | Argoverse 2 val · Argoverse 2 val | minADE6: 0.901 -> 0.713 | Helped |
| use 6 instead of 4 mixture components (ablation) | SeNeVA · INTERACTION val | minFDE6: 0.5362 -> 0.4306 | Helped |
| use 6 instead of 4 mixture components (ablation) | SeNeVA · INTERACTION val | minADE6: 0.4418 -> 0.1967 | Helped |
| use 6 instead of 4 mixture components (ablation) | SeNeVA · INTERACTION val | MR: 0.1065 -> 0.0790 | Helped |
| increase mixture components from 6 to 16 (ablation) | SeNeVA · INTERACTION val | MR: 0.0790 -> 0.1280 | Hurt |
| include the assignment network and NMS sampling versus means-based sampling (ablation) | SeNeVA · INTERACTION val | minADE6: 0.2281 -> 0.1967 | Helped |
Surprising findings
- Predictive entropy for OOD intersection cases is 0.5% lower than for in-distribution cases, contrary to the expected trend.
Limitations
- Requires target-centric inputs and predicts for only a single agent at a time.
- Multi-agent scenarios need parallel inference of several models, which is computationally expensive.
- OOD entropy separation is small in some settings (intersection case reversed).
Experimental setup
Datasets: INTERACTION Argoverse 2
Benchmarks: INTERACTION Argoverse 2
Metrics: minFDE6 minADE6 MR6 predictive entropy
Baselines: DenseTNT Multi-Branch SS-ASP MultiModalTransformer HDGT DESIRE MultiPath TNT Forecast-MAE Deep Ensembles
Builds on: VectorNet LaneGCN variational autoencoder Deep Ensembles
Closest papers by meaning
- SmartRefine: A Scenario-Adaptive Refinement Framework for Efficient Motion Prediction (CVPR 2024 · similarity 0.85)
- SmartPretrain: Model-Agnostic and Dataset-Agnostic Representation Learning for Motion Prediction (ICLR 2025 · similarity 0.84)
- CUQDS: Conformal Uncertainty Quantification Under Distribution Shift for Trajectory Prediction (AAAI 2025 · similarity 0.84)
- CaDeT: a Causal Disentanglement Approach for Robust Trajectory Prediction in Autonomous Driving (CVPR 2024 · similarity 0.84)
- Proper Scoring Rules for Agentic Uncertainty Quantification (NeurIPS 2026 · similarity 0.84)
- DeMo: Decoupling Motion Forecasting into Directional Intentions and Dynamic States (NeurIPS 2024 · similarity 0.84)
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.