Research map / Detection and segmentation / CNN features
MGMapNet: Multi-Granularity Representation Learning for End-to-End Vectorized 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 MGMapNet, which represents vectorized HD map elements with both instance-level and point-level queries aggregated from multi-scale BEV features, linked by point-instance interaction attention, reaching SOTA single-frame mAP on nuScenes and Argoverse2.
Problem
Vectorized HD map construction methods use either point-level queries (good geometry, may miss elements) or instance-level queries (good category, imprecise geometry), and do not exploit the relationship between the two granularities.
Approach
A Multi-Granularity Aggregator generates instance-level queries (aggregating over the whole instance scope) and point-level queries (aggregating locally) from multi-scale BEV features via deformable sampling, and a Point Instance Interaction module (P2P and P2I attention) exchanges information between the two granularities within each decoder layer.
Key results
- On nuScenes val at 24 epochs, MGMapNet reaches 66.8 mAP, exceeding MapTRv2 (61.5) by 5.3 mAP.
- At 110 epochs on nuScenes, MGMapNet reaches 73.6 mAP vs. 68.7 for MapTRv2 and 72.6 for MapQR.
- On Argoverse2 val (6 epochs), MGMapNet reaches 71.2 mAP (2D) and 69.1 mAP (3D), improving over MapTRv2 by 3.8 and 4.4.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| use MGMapNet instead of MapQR | nuScenes val, 24 epochs · nuScenes val, 24 epochs | mAP: 66.4 -> 66.8 | Helped |
| use MGMapNet instead of MapQR | nuScenes val, 110 epochs · nuScenes val, 110 epochs | mAP: 72.6 -> 73.6 | Helped |
| use MGMapNet instead of HIMap | Argoverse2 val, 2D, 6 epochs · Argoverse2 val, 2D, 6 epochs | mAP: 69.6 -> 71.2 | Helped |
| use MGMapNet instead of HIMap | Argoverse2 val, 3D, 6 epochs · Argoverse2 val, 3D, 6 epochs | mAP: 68.4 -> 69.1 | Helped |
| replace Multi-Point Attention with Multi-Granularity Attention (ablation) | MGMapNet · nuScenes val, 24 epochs | mAP: 55.9 -> 63.6 (+7.7) | Helped |
| add P2P attention to Multi-Granularity Aggregator (ablation) | MGMapNet · nuScenes val, 24 epochs | mAP: 62.7 -> 64.8 | Helped |
| add P2I attention to Multi-Granularity Aggregator (ablation) | MGMapNet · nuScenes val, 24 epochs | mAP: 62.7 -> 65.0 | Helped |
| add auxiliary loss (ablation) | MGMapNet · nuScenes val, 24 epochs | mAP: +0.8 | Helped |
| add multi-scale BEV features (ablation) | MGMapNet · nuScenes val, 24 epochs | mAP: +0.6 | Helped |
| add reference point positional encoding (ablation) | MGMapNet · nuScenes val, 24 epochs | mAP: +1.2 | Helped |
| increase query number to 125 from 100 (ablation) | MGMapNet · nuScenes val, 24 epochs | mAP: 66.8 -> 66.4 | Hurt |
| set repeat number Nrep to 4 instead of 8 (ablation) | MGMapNet · nuScenes val, 24 epochs | mAP: 66.8 -> 63.7 | Hurt |
Surprising findings
- Ablation shows a P2P-only or P2I-only variant each gives a similar gain (64.8 and 65.0), with the combination reaching 66.8.
Limitations
- Real-time performance is not addressed; FPS (11.7) is below MapTRv2 (14.1).
- Only single-frame results; temporal information is not exploited.
- Argoverse2 long-training results are reported only for 2D configuration.
Experimental setup
Datasets: nuScenes Argoverse2
Benchmarks: nuScenes Argoverse2
Metrics: mAP (Chamfer distance) IoU-based AP FPS parameters
Baselines: HDMapNet BeMapNet PivotNet MapTRv2 MGMap MapQR VectorMapNet MapVR HIMap StreamMapNet
Builds on: MapTRv2 StreamMapNet Deformable DETR LSS
Closest papers by meaning
- HIMap: HybrId Representation Learning for End-to-end Vectorized HD Map Construction (CVPR 2024 · similarity 0.86)
- MGMap: Mask-Guided Learning for Online Vectorized HD Map Construction (CVPR 2024 · similarity 0.85)
- InteractionMap: Improving Online Vectorized HDMap Construction with Interaction (CVPR 2025 · similarity 0.83)
- MM-Point: Multi-View Information-Enhanced Multi-Modal Self-Supervised 3D Point Cloud Understanding (AAAI 2024 · similarity 0.83)
- Uncertainty-Instructed Structure Injection for Generalizable HD Map Construction (CVPR 2025 · similarity 0.82)
- Multi-view Granular-ball Contrastive Clustering (AAAI 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.