Research map / Detection and segmentation / CNN features
SDTagNet: Leveraging Text-Annotated Navigation Maps for Online 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 SDTagNet, which encodes OpenStreetMap points, polylines and relations with contrastively pretrained BERT tag embeddings and ORF element identifiers, feeding an SD map prior into MapTRv2 via cross-attention for far-range online HD map construction.
Problem
Online HD map construction is limited by onboard sensor range; existing SD map prior methods use only hand-selected road classes and ignore points, relations and free-text OSM tags.
Approach
Encode all OSM element types at point level with a self-supervised, contrastively pretrained BERT tag embedding and orthogonal random feature element identifiers, so open-vocabulary textual annotations and points/polylines/relations are used jointly without manual feature selection.
Key results
- On Argoverse 2 far range (120m x 60m), SDTagNet reaches 18.9 mAP vs 13.0 for MapTRv2 without prior (+5.9 mAP, +45%) and +3.2 mAP over the best SD-prior baseline.
- On nuScenes far range, SDTagNet reaches 8.0 mAP vs 3.9 for MapTRv2 (+4.1 mAP, +105%).
- Under the Argoverse 2 original split with the MapTracker-style label setting, SDTagNet achieves 78.0 mAP near range vs 76.9 for MapTracker.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Add SDTagNet SD map priors | MapTRv2 · Argoverse 2 far range, geographic split | mAP: 13.0 -> 18.9; +5.9 | Helped |
| Add SDTagNet SD map priors | MapTRv2 · nuScenes far range, geographic split | mAP: 3.9 -> 8.0; +4.1 | Helped |
| Add SDTagNet SD map priors | MapTRv2 · Argoverse 2 near range, geographic split | mAP: 46.5 -> 48.1 | Helped |
| Add SDTagNet SD map priors | MapTRv2 · nuScenes near range, geographic split | mAP: 22.9 -> 23.3 | No clear effect |
| Add point-level queries without ORF identifiers (ablation) | SMERF† all-info encoder · Argoverse 2 near range | mAP: 45.9 -> 44.3 | Hurt |
| Add ORF identifiers alongside point-level queries (ablation) | SMERF† all-info encoder · Argoverse 2 near range | mAP: 44.3 -> 46.3 | Helped |
| Add NLP tag embeddings alone (ablation) | SMERF† all-info encoder · Argoverse 2 near range | mAP: 45.9 -> 45.8 | No clear effect |
| Add point-level queries, NLP embeddings, and ORF identifiers (ablation) | SMERF† all-info encoder · Argoverse 2 near range | mAP: 45.9 -> 48.1 | Helped |
| Add BEV features to SDTagNet cross-attention (ablation) | SDTagNet · Argoverse 2 near range | mAP: 48.1 -> 48.1 | No clear effect |
| Fine-tune the NLP encoder at learning-rate multiplier × 0.1 rather than freezing it (ablation) | NLP tag embedding module · Argoverse 2 near range | mAP: 45.6 -> 48.1 | Helped |
| Use relevant-tag contrastive pretraining (ablation) | NLP tag embedding module · Argoverse 2 near range | mAP: 47.5 -> 48.1 | Helped |
| Use 8 attention heads with dropout 0.1 rather than 4 heads or dropout 0.2 (ablation) | SD map encoder cross-attention · Argoverse 2 near range | mAP: 46.9 -> 48.1; 47.7 -> 48.1 | Helped |
Surprising findings
- Near-range performance gains are negligible or negative for all SD map prior methods on Argoverse 2 and nuScenes, attributed to geographic train/val overlap and weaker base architectures.
- Data augmentation of SD map priors (position noise, element dropping, tag masking) did not improve validation performance.
Limitations
- Assumes congruency between OSM SD map and HD map ground truth; discrepancies (e.g., missing tunnels, service roads in nuScenes) hurt accuracy.
- One-way driving direction is sometimes predicted wrongly because point order is not encoded in element identifiers.
- Near-range gains are absent; SD priors may aggravate geographic overfitting.
Experimental setup
Datasets: OpenStreetMap Argoverse 2 nuScenes
Benchmarks: Argoverse 2 nuScenes
Metrics: mAP AP per class (divider, solid, boundary, centerline, pedestri FPS parameter count VRAM
Baselines: MapTRv2 PMapNet SMERF MapTracker StreamMapNet HIMap MapUnveiler
Builds on: MapTRv2 SMERF PMapNet BERT orthogonal random features
Closest papers by meaning
- Leveraging SD Map to Augment HD Map-based Trajectory Prediction (CVPR 2025 · similarity 0.89)
- MapExpert: Online HD Map Construction with Simple and Efficient Sparse Map Element Expert (AAAI 2025 · similarity 0.88)
- InteractionMap: Improving Online Vectorized HDMap Construction with Interaction (CVPR 2025 · similarity 0.87)
- OpenSatMap: A Fine-grained High-resolution Satellite Dataset for Large-scale Map Construction (NeurIPS 2024 · similarity 0.87)
- Producing and Leveraging Online Map Uncertainty in Trajectory Prediction (CVPR 2024 · similarity 0.86)
- Map It Anywhere: Empowering BEV Map Prediction using Large-scale Public Datasets (NeurIPS 2024 · similarity 0.86)
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.