Research map / Label-efficient and robust learning / Recognition and long-tail
EDA: Evolving and Distinct Anchors for Multimodal 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 Evolving and Distinct Anchors (EDA), which updates anchors with intermediate decoder outputs and applies NMS to select distinct anchors before matching, improving MTR for multimodal motion prediction on Waymo Open Motion Dataset.
Problem
Positive component selection in mixture-model motion prediction: prediction-based matching yields clustered trajectories that are hard to score, while anchor-based matching has limited regression capability due to static, sparse anchors.
Approach
Anchors are predefined and then replaced by intermediate decoder-layer predictions (evolving anchors), and only NMS-selected distinct anchors are labeled positive or negative before matching to ground truth, with the rest treated as neutral.
Key results
- On the WOMD test set, EDA reaches mAP 0.4401 and minADE 0.5718 versus MTR's 0.4129 and 0.6050.
- EDA achieves a 13.5% relative reduction in Miss Rate over MTR (0.1169 vs 0.1351 on test).
- Distinct anchors with twice-evolving anchors give the best top-6 metrics on validation (Miss Rate 0.1178 vs 0.1348 for the CE baseline).
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Apply EDA to anchor-based QCNet | Anchor QCNet · Argoverse 2 validation | mAP: 0.48 -> 0.58 | Helped |
| Use twice-evolving distinct anchors instead of static anchors (ablation) | MTR · Waymo Open Motion Dataset validation | Miss Rate: 0.1348 -> 0.1178 | Helped |
| Increase anchor updates from 0 to 1 (ablation) | MTR · Waymo Open Motion Dataset validation, top 6 | minADE: 0.6012 -> 0.5867 | Helped |
| Increase anchor updates from 1 to 2 without distinct anchors (ablation) | EDA with BCE · Waymo Open Motion Dataset validation, top 6 | rank mAP: 0.4228 -> 0.4207 | Hurt |
| Replace CE with BCE (ablation) | EDA with 1 evolving time · Waymo Open Motion Dataset validation, top 6 | rank mAP: 0.4183 -> 0.4228 | Helped |
| Add distinct anchors (ablation) | EDA with 1 evolving time and BCE · Waymo Open Motion Dataset validation, top 6 | rank mAP: 0.4228 -> 0.4278 | Helped |
| Add distinct anchors (ablation) | EDA with 2 evolving times and BCE · Waymo Open Motion Dataset validation, top 6 | rank mAP: 0.4207 -> 0.4353 | Helped |
| Add distinct anchors (ablation) | MTR with static anchors and BCE · Waymo Open Motion Dataset validation, top 6 | mAP original: 0.4171 -> 0.4208 | No clear effect |
| Increase anchor updates to 5 and add distinct anchors (ablation) | EDA · Waymo Open Motion Dataset validation, top 6 | Miss Rate: 0.1178 -> 0.1245 | Hurt |
| Use 5 evolving times with distinct anchors instead of 2 (ablation) | EDA with BCE · Waymo Open Motion Dataset validation, top 6 | rank mAP: 0.4353 -> 0.4322 | Hurt |
| Use 16 anchors and increase evolving times from 0 to 2 with distinct anchors (ablation) | EDA · Waymo Open Motion Dataset validation | mAP: 0.4009 -> 0.4207 | Helped |
| Use 100 anchors and increase evolving times from 0 to 1 with distinct anchors (ablation) | EDA · Waymo Open Motion Dataset validation | mAP: 0.4158 -> 0.4302 | Helped |
Surprising findings
- Using only distinct anchors on static anchors brings little to no improvement, so the gains come from combining evolving and distinct anchors.
- Five anchor updates performed worse than twice-evolving anchors even with distinct anchors, indicating too many updates hurt.
Limitations
- Gains diminish or reverse with too many anchor evolutions due to prediction clustering.
- Distinct anchors alone give negligible benefit without evolving anchors.
Experimental setup
Datasets: Waymo Open Motion Dataset Argoverse 2
Benchmarks: Waymo Open Motion Dataset
Metrics: minADE minFDE Miss Rate Overlap Rate mAP Soft mAP
Baselines: MTR MTR++ MotionCNN ReCoAt DenseTNT SceneTransformer HDGT QCNet
Builds on: MTR Dense Distinct Query (DDQ) Gaussian mixture model winner-takes-all loss
Closest papers by meaning
- SmartPretrain: Model-Agnostic and Dataset-Agnostic Representation Learning for Motion Prediction (ICLR 2025 · similarity 0.86)
- ModeSeq: Taming Sparse Multimodal Motion Prediction with Sequential Mode Modeling (CVPR 2025 · similarity 0.86)
- UniMotion: A Unified Motion Framework for Simulation, Prediction and Planning (NeurIPS 2025 · similarity 0.85)
- SingularTrajectory: Universal Trajectory Predictor Using Diffusion Model (CVPR 2024 · similarity 0.84)
- MotionMap: Representing Multimodality in Human Pose Forecasting (CVPR 2025 · similarity 0.84)
- Towards Robust Multimodal Open-set Test-time Adaptation via Adaptive Entropy-aware Optimization (ICLR 2025 · 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.