Research map / 3D vision / 3D generation
Accelerating Online Mapping and Behavior Prediction via Direct BEV Feature Attention
Xunjiang Gu, Guanyu Song, Igor Gilitschenski, Marco Pavone, Boris Ivanovic
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| feed BEV features directly to the prediction model instead of decoded vector maps | HiVT with MapTR · nuScenes | minFDE: 0.8900 -> 0.7401 | Improved |
| feed BEV features to DenseTNT instead of vector maps | DenseTNT with MapTR · nuScenes | minFDE: 2.0661 -> 1.4700 | Improved |
| add BEV features when the map model also outputs centerlines | MapTRv2-CL with HiVT/DenseTNT · nuScenes | minFDE: DenseTNT 1.3174 -> 1.3609 | No clear effect |
| add uncertainty-enhanced vertices (prior work) for DenseTNT with MapTR | DenseTNT · nuScenes | minFDE: 2.0661 -> 2.1502 | Hurt |
| skip map decoding and attend to BEV features directly | HiVT with MapTR · nuScenes inference | runtime (ms): 22.4 -> 9.1 | Improved |
| skip map decoding with StreamMapNet | HiVT with StreamMapNet · nuScenes inference | runtime (ms): 33.6 -> 29.4 | Improved |
| use BEV patch 10x5 instead of 20x20 (ablation) | HiVT + MapTRv2-Centerline · nuScenes | minFDE: 0.7518 -> 0.8003 | Hurt |
| use larger BEV patch 40x20 instead of 20x20 (ablation) | HiVT + MapTRv2-Centerline · nuScenes | MR: 0.0737 -> 0.0770 | Hurt |
| use temporal BEV features (BEVFormer-based MapTR, StreamMapNet) instead of single-frame LSS features (ablation) | BEV feature prediction · nuScenes | minADE/minFDE/MR: 16%, 19%, 24% vs 4%, 2%, 4% relative improvement | Improved |
Closest papers by meaning
- Producing and Leveraging Online Map Uncertainty in Trajectory Prediction (CVPR 2024 · similarity 0.91)
- MapExpert: Online HD Map Construction with Simple and Efficient Sparse Map Element Expert (AAAI 2025 · similarity 0.91)
- Map It Anywhere: Empowering BEV Map Prediction using Large-scale Public Datasets (NeurIPS 2024 · similarity 0.89)
- Leveraging SD Map to Augment HD Map-based Trajectory Prediction (CVPR 2025 · similarity 0.88)
- SDTagNet: Leveraging Text-Annotated Navigation Maps for Online HD Map Construction (NeurIPS 2025 · similarity 0.87)
- Unveiling the Hidden: Online Vectorized HD Map Construction with Clip-Level Token Interaction and Propagation (NeurIPS 2024 · similarity 0.87)
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.
