Research map / Probabilistic and scientific modeling / Molecules and proteins
Uncertainty-Instructed Structure Injection for Generalizable HD Map Construction
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 UIGenMap, an uncertainty-instructed PV detection branch whose probabilistic prompts are injected into a BEV HD-map decoder and distilled into mimic queries, evaluated on geo-disjoint nuScenes and Argoverse2 splits.
Problem
Online HD map vectorization generalizes poorly to unfamiliar driving scenes because training and validation splits share geographic overlap and BEV transformation introduces geometric errors and texture loss.
Approach
Uses a pre-trained perspective-view (PV) detection branch with uncertainty-aware decoding (Gaussian-resampled attention and Laplace point uncertainty) to build uncertainty-weighted PV prompts, injects them into BEV features and map queries via hybrid injection, and distills them into mimic queries so PV branch is not needed at inference.
Key results
- On nuScenes region-based split with ResNet-50, UIGenMap reaches 39.8 mAP (+5.7 over StreamMapNet's 34.1).
- On nuScenes city-based split, UIGenMap achieves 23.6 mAP (+4.3 over StreamMapNet's 19.3).
- The lightweight UIGenMap-d variant using mimic queries reaches 39.3 mAP region-based at 12.2 FPS versus 8.2 FPS for the full model.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use UIGenMap instead of StreamMapNet | HD map construction · nuScenes region-based | mAP: 34.1 -> 39.8 | Helped |
| Use UIGenMap instead of MapTRv2 | HD map construction · nuScenes city-based | mAP: 21.8 -> 23.6 | Helped |
| Use UIGenMap instead of StreamMapNet | HD map construction · Argoverse2 region-based | mAP: 58.3 -> 60.4 | Helped |
| Use UIGenMap instead of StreamMapNet | HD map construction · Argoverse2 city-based | mAP: 42.3 -> 44.8 | Helped |
| Use UIGenMap instead of SQD-MapNet | HD map construction · nuScenes original partition | mAP: 63.9 -> 64.8 | Helped |
| Use UIGenMap instead of SQD-MapNet | HD map construction · Argoverse2 original partition | mAP: 63.3 -> 63.8 | Helped |
| Add the shared PV image backbone (ablation) | nuScenes region-based baseline · nuScenes region-based | mAP: 34.1 -> 37.4 | Helped |
| Add UI2DPrompt (ablation) | nuScenes region-based baseline · nuScenes region-based | mAP: 34.1 -> 37.1 | Helped |
| Add UA-Decoder (ablation) | nuScenes region-based UIGenMap · nuScenes region-based | mAP: 38.7 -> 39.8 | Helped |
| Add UA-Attention (ablation) | nuScenes region-based baseline decoder · nuScenes region-based | mAP: 34.1 -> 35.0 | Helped |
| Add UA-Head (ablation) | nuScenes region-based baseline decoder · nuScenes region-based | mAP: 34.1 -> 36.3 | Helped |
| Use IPM instead of depth-based 3D transformation for PV coordinates (ablation) | UI2DPrompt with uncertainty weight and embedding · nuScenes region-based | mAP: 38.7 -> 39.8 | Helped |
Limitations
- Methods are designed and evaluated primarily on specific datasets (nuScenes and Argoverse2).
- Full model runs at only 8.2 FPS, and the mimic-query variant is needed for real-time speed.
Experimental setup
Datasets: nuScenes Argoverse2
Benchmarks: nuScenes region-based split nuScenes city-based split Argoverse2 region-based split Argoverse2 city-based split nuScenes original split Argoverse2 original split
Metrics: mAP AP per class (ped crossing, divider, boundary) FPS
Baselines: VectorMapNet MapTR MGMap MapQR GeMap MapTRv2 StreamMapNet SQD-MapNet
Builds on: StreamMapNet MapTR BEVFormer Deformable attention Laplace uncertainty loss from StreamMapNet/uncertainty vector map work
Closest papers by meaning
- SafeMap: Robust HD Map Construction from Incomplete Observations (ICML 2025 · similarity 0.88)
- MapExpert: Online HD Map Construction with Simple and Efficient Sparse Map Element Expert (AAAI 2025 · similarity 0.87)
- InteractionMap: Improving Online Vectorized HDMap Construction with Interaction (CVPR 2025 · similarity 0.87)
- MGMap: Mask-Guided Learning for Online Vectorized HD Map Construction (CVPR 2024 · similarity 0.86)
- SDTagNet: Leveraging Text-Annotated Navigation Maps for Online HD Map Construction (NeurIPS 2025 · similarity 0.86)
- Compact HD Map Construction via Douglas-Peucker Point Transformer (AAAI 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.