Research map / 3D vision / 3D generation
SafeMap: Robust HD Map Construction from Incomplete Observations
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 SafeMap, a plug-and-play framework that makes camera-based online HD map construction robust to missing camera views via Gaussian-sampled perspective-view reconstruction and distillation-based BEV correction.
Problem
Online vectorized HD map construction degrades severely when one or more multi-view camera images are missing or corrupted, and prior robustness work either assumes complete views or targets 3D detection.
Approach
Reconstruct a missing camera view by a learnable query that deformably attends to all available views tiled into a panoramic perspective view, with Gaussian-distributed reference points centered on the missing view's position, and correct incomplete BEV features by MSE distillation from panoramic complete-view BEV features.
Key results
- On nuScenes with MapTR, SafeMap raises complete-view mAP from 50.3 to 52.5 and incomplete-view mAP from 39.9 to 47.3 (average over six single-view-missing cases).
- On nuScenes with HIMap, SafeMap improves mAP by 2.4% to 18.2% across missing-view cases, and average incomplete mAP from 52.1 to 60.1.
- On Argoverse2 with MapTR, SafeMap improves complete-view mAP from 58.7 to 59.7 and front-view-missing mAP from 51.5 to 55.6.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| add SafeMap modules to MapTR | MapTR · nuScenes val, all views | mAP: 50.3 -> 52.5 | Helped |
| add SafeMap modules to HIMap | HIMap · nuScenes val, all views | mAP: 65.5 -> 66.0 | Helped |
| add SafeMap modules to MapTR | MapTR · nuScenes val, front view missing | mAP: 31.3 -> 42.4 | Helped |
| add SafeMap modules to HIMap | HIMap · nuScenes val, front view missing | mAP: 38.0 -> 54.3 | Helped |
| add SafeMap modules to MapTR | MapTR · Argoverse2 val, all views | mAP: 58.7 -> 59.7 | Helped |
| add SafeMap modules to MapTR | MapTR · Argoverse2 val, front view missing | mAP: 51.5 -> 55.6 | Helped |
| add G-PVR and D-BEVC (ablation) | SafeMap (MapTR) · nuScenes val, six missing-view cases average | mAP: 39.9 -> 47.3 | Helped |
| add G-PVR (ablation) | SafeMap (MapTR) · nuScenes val, six missing-view cases average | mAP: 39.9 -> 46.5 | Helped |
| add D-BEVC (ablation) | SafeMap (MapTR) · nuScenes val, six missing-view cases average | mAP: 39.9 -> 46.5 | Helped |
| replace G-PVR with Mean-PVR (ablation) | SafeMap (MapTR) · nuScenes val, six missing-view cases average | mAP: 47.3 -> 39.9 | Hurt |
| replace G-PVR with MAE-PVR (ablation) | SafeMap (MapTR) · nuScenes val, six missing-view cases average | mAP: 47.3 -> 46.5 | Hurt |
| replace G-PVR with Standard-PVR (ablation) | SafeMap (MapTR) · nuScenes val, six missing-view cases average | mAP: 47.3 -> 46.8 | Hurt |
Surprising findings
- The Mean-PVR variant underperforms the no-reconstruction baseline (39.9 vs 46.5 mAP in ablation table), indicating naive view fusion hurts.
Limitations
- Does not address multi-sensor fusion (e.g., LiDAR) when camera data is incomplete, per the Impact Statement.
- Robustness evaluated only on camera-view dropout and MapBench corruptions; gains on Argoverse2 are smaller because of seven cameras.
- Models are trained with random view masking, and the paper does not evaluate under unseen failure types beyond the tested cases.
Experimental setup
Datasets: nuScenes Argoverse2
Benchmarks: nuScenes Argoverse2 MapBench corruptions
Metrics: AP mAP mRR mCE FPS GPU memory parameters
Baselines: MapTR HIMap M-BEV (local PVR) MetaBEV UniBEV
Builds on: MapTR HIMap deformable attention M-BEV
Closest papers by meaning
- Uncertainty-Instructed Structure Injection for Generalizable HD Map Construction (CVPR 2025 · similarity 0.88)
- MapExpert: Online HD Map Construction with Simple and Efficient Sparse Map Element Expert (AAAI 2025 · similarity 0.86)
- Is Your HD Map Constructor Reliable under Sensor Corruptions? (NeurIPS 2024 · similarity 0.86)
- Map It Anywhere: Empowering BEV Map Prediction using Large-scale Public Datasets (NeurIPS 2024 · similarity 0.85)
- Unveiling the Hidden: Online Vectorized HD Map Construction with Clip-Level Token Interaction and Propagation (NeurIPS 2024 · similarity 0.84)
- M-BEV: Masked BEV Perception for Robust Autonomous Driving (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.