Research map / Detection and segmentation / CNN features
MGMap: Mask-Guided Learning for Online 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 MGMap, which uses learned instance masks to activate lane queries (MAI decoder) and binary masks to extract point-level patch features for refinement (PG-MPR), improving online vectorized HD map construction by about 10 mAP over MapTR.
Problem
Online vectorized HD map construction from onboard sensors struggles with sparse, subtle map annotations and strong shape priors, causing ambiguous feature localization and loss of detailed structures in detection-based frameworks like MapTR.
Approach
Uses learned instance segmentation masks to dynamically initialize lane queries from enhanced multi-level BEV features, and binary mask features with position-guided ROIAlign patches to refine point positions, replacing fixed query designs and sparse deformable sampling.
Key results
- On nuScenes val (camera, R50, 30 epochs, 60m x 30m), MGMap reaches 61.4 mAP (APchamfer) vs 51.1 for MapTR, a 10.3 mAP gain.
- LiDAR-only (SECOND) reaches 67.9 mAP and camera-LiDAR fusion reaches 71.7 mAP on nuScenes val.
- On Argoverse2 subset, MGMap achieves 62.8 mAP vs 57.4 for MapTR.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use mask-guided MGMap instead of MapTR | Online HD map construction with ResNet50 · nuScenes val, camera, 60m × 30m | APchamfer (avg): 51.1 -> 61.4 | Helped |
| Use MGMap instead of MapTR | Online HD map construction with ResNet50 · Argoverse2 subset | mAP: 57.4 -> 62.8 | Helped |
| Add instance-level MAI (ablation) | MGMap without mask guidance · nuScenes val, camera, 60m × 30m | mAP: 57.6 -> 59.5 | Helped |
| Add point-level PG-MPR (ablation) | MGMap without mask guidance · nuScenes val, camera, 60m × 30m | mAP: 57.6 -> 60.2 | Helped |
| Add both MAI and PG-MPR (ablation) | MGMap without mask guidance · nuScenes val, camera, 60m × 30m | mAP: 57.6 -> 61.4 | Helped |
| Use a multi-level neck in PV space rather than no multi-level neck (ablation) | MGMap baseline · nuScenes val, camera, 60m × 30m | mAP: 51.1 -> 48.1 | Hurt |
| Add a multi-level EML neck in BEV space (ablation) | MGMap baseline · nuScenes val, camera, 60m × 30m | mAP: 51.1 -> 52.2 | Helped |
| Set PG-MPR patch size to 0.08 with two refinement stages (ablation) | MGMap · nuScenes val, camera, 60m × 30m | mAP: 59.5 -> 60.8 | Helped |
| Set PG-MPR patch size to 0.1 and use one refinement stage (ablation) | MGMap · nuScenes val, camera, 60m × 30m | mAP: 59.5 -> 60.5 | Helped |
| Use three PG-MPR refinement stages instead of two (ablation) | MGMap with d=0.1 · nuScenes val, camera, 60m × 30m | mAP: 61.4 -> 60.6 | Hurt |
| Increase PG-MPR patch size from 0.1 to 0.12 (ablation) | MGMap with two refinement stages · nuScenes val, camera, 60m × 30m | mAP: 61.4 -> 60.1 | Hurt |
Surprising findings
- Putting the PV-space FPN instead of BEV-space neck degrades performance (48.1 mAP vs 51.1 baseline), while the BEV-level EML neck helps.
Limitations
- Inference speed is lower than MapTR (11.6 vs 15.7 FPS for camera-based).
- Patch size and refinement stages require hyperparameter tuning (d=0.1, s=2 chosen).
Experimental setup
Datasets: nuScenes Argoverse2
Benchmarks: nuScenes Argoverse2
Metrics: APchamfer APraster mAP FPS
Baselines: MapTR HDMapNet VectorMapNet MapVR BeMapNet PivotNet InstaGraM MapTRV2 MapMaskFormer
Builds on: MapTR deformable DETR BEVFormer MapVR
Closest papers by meaning
- Unveiling the Hidden: Online Vectorized HD Map Construction with Clip-Level Token Interaction and Propagation (NeurIPS 2024 · similarity 0.86)
- Uncertainty-Instructed Structure Injection for Generalizable HD Map Construction (CVPR 2025 · similarity 0.86)
- MapExpert: Online HD Map Construction with Simple and Efficient Sparse Map Element Expert (AAAI 2025 · similarity 0.86)
- MGMapNet: Multi-Granularity Representation Learning for End-to-End Vectorized HD Map Construction (ICLR 2025 · similarity 0.85)
- HIMap: HybrId Representation Learning for End-to-end Vectorized HD Map Construction (CVPR 2024 · similarity 0.85)
- Compact HD Map Construction via Douglas-Peucker Point Transformer (AAAI 2024 · similarity 0.85)
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.