Research map / Vision-language and multimodal / Multimodal LLMs
Is Your HD Map Constructor Reliable under Sensor Corruptions?
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.
Introduces MapBench, a benchmark of 29 sensor corruption types (camera, LiDAR, and camera-LiDAR combinations) on nuScenes to evaluate robustness of HD map constructors, reporting large degradation under snow and sensor failures and identifying augmentation and architecture factors that help.
Problem
HD map constructors perform well on clean data, but their robustness to adverse weather and sensor failures under real-world corruptions is unknown, raising safety concerns.
Approach
A corruption benchmark for vectorized HD map construction covering camera-only, LiDAR-only, and camera-LiDAR fusion settings, with three severity levels and relative robustness metrics (mCE and mRR) normalized to MapTR baselines.
Key results
- Under camera Frame Lost, MapTR fusion mAP drops by 26.2% and HIMap by 48.1%; camera unavailability drops MapTR and HIMap mAP by 40.0% and 68.9%.
- Snow and Cross-Sensor LiDAR corruptions cause more than 80% performance drops for all LiDAR-only methods.
- Swin Transformer backbone improves mCE of PivotNet and BeMapNet by 22.2% and 24.1% absolute versus ResNet-50.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use SwinT instead of ResNet-50 (ablation) | PivotNet · Camera-only corruption benchmark | mCE: 100.0 -> 77.8 | Helped |
| Use SwinT instead of ResNet-50 (ablation) | BeMapNet · Camera-only corruption benchmark | mCE: 100.0 -> 75.9 | Helped |
| Use BEVPool instead of GKT (ablation) | MapTR · Camera-only corruption benchmark | mRR: 49.3 -> 48.1 | Hurt |
| Add temporal fusion (ablation) | StreamMapNet · Corruption benchmark | mRR: 47.1 -> 55.5 | Helped |
| Add temporal fusion (ablation) | StreamMapNet · Corruption benchmark | mCE: 100.0 -> 85.9 | Helped |
| Train for 110 instead of 24 epochs (ablation) | MapTR · Camera-only corruption benchmark | mCE: 100.0 -> 80.9 | Helped |
| Train for 110 instead of 30 epochs (ablation) | PivotNet · Camera-only corruption benchmark | mCE: 100.0 -> 90.2 | Helped |
| Train for 110 instead of 30 epochs (ablation) | BeMapNet · Camera-only corruption benchmark | mCE: 100.0 -> 89.2 | Helped |
| Add PhotoMetric augmentation (ablation) | MapTR-R50 · Camera corruption benchmark | mRR: 41.1 -> 49.3 | Helped |
| Add PhotoMetric augmentation (ablation) | MapTR-R50 · Camera corruption benchmark | mCE: 100.0 -> 84.5 | Helped |
| Add Rotate augmentation (ablation) | MapTR-R50 · Camera corruption benchmark | mRR: 41.1 -> 38.1 | Hurt |
| Add Flip augmentation (ablation) | MapTR-R50 · Camera corruption benchmark | mRR: 41.1 -> 38.7 | Hurt |
Surprising findings
- HIMap has higher clean accuracy than MapTR in fusion but its robustness under camera corruption is relatively poorer.
- Incomplete Echo causes negligible drops for most LiDAR-only models.
Limitations
- Benchmark cannot cover the full range of real-world out-of-distribution contexts.
- Temporally aggregated LiDAR corruption does not ensure temporal consistency.
- Only standard augmentations explored; advanced multi-modal augmentation left open.
Experimental setup
Datasets: nuScenes
Benchmarks: MapBench (nuScenes val corruptions)
Metrics: mAP mCE mRR AP per class
Baselines: HDMapNet VectorMapNet PivotNet BeMapNet MapTR MapTRv2 StreamMapNet HIMap
Builds on: RoboBEV Robo3D RoboDepth MapTR nuScenes
Closest papers by meaning
- SafeMap: Robust HD Map Construction from Incomplete Observations (ICML 2025 · similarity 0.86)
- Unveiling the Hidden: Online Vectorized HD Map Construction with Clip-Level Token Interaction and Propagation (NeurIPS 2024 · similarity 0.83)
- Compact HD Map Construction via Douglas-Peucker Point Transformer (AAAI 2024 · similarity 0.82)
- SDTagNet: Leveraging Text-Annotated Navigation Maps for Online HD Map Construction (NeurIPS 2025 · similarity 0.82)
- The Road Ahead in Autonomous Driving: The KITScenes Multimodal Dataset (NeurIPS 2026 · similarity 0.81)
- Driving by the Rules: A Benchmark for Integrating Traffic Sign Regulations into Vectorized HD Map (CVPR 2025 · similarity 0.81)
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.