Research map / Detection and segmentation / CNN features
InteractionMap: Improving Online Vectorized HDMap Construction with Interaction
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 InteractionMap, a DETR-like online vectorized HD map construction method that adds point- and instance-level relation embeddings, key-frame-based hierarchical temporal fusion, and geometry-aware focal loss/cost, reaching SOTA mAP on nuScenes and Argoverse2.
Problem
Online vectorized HD map construction from multi-view cameras suffers from weak instance-level shape modeling, temporal inconsistency across frames under occlusion, and misalignment between classification confidence and localization quality.
Approach
Combines explicit point-wise and instance-wise relation embeddings injected into decoder self-attention, a key-frame-based local-to-global streaming/stacking temporal fusion, and geometry-aware classification scores (point-to-point, edge direction, GIoU) used in both the focal loss and the matching cost.
Key results
- On nuScenes at 110 epochs, InteractionMap-C reaches 76.0 mAP and InteractionMap-R 75.5 mAP, versus 68.7 for MapTRv2 and 72.6 for MapQR.
- On nuScenes at 24 epochs, InteractionMap-C reaches 71.8 mAP and InteractionMap-R 71.6 mAP, versus 61.5 for MapTRv2.
- Ablation on nuScenes shows relation embedding adds 4.5 mAP, geometry alignment adds 7.5 mAP, and key-frame temporal fusion adds about 5 mAP over the baseline.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Add relation embedding module (ablation) | MapTR baseline · nuScenes val | mAP: +4.5 mAP | Helped |
| Add geometry-aware alignment module (ablation) | MapTR baseline · nuScenes val | mAP: 7.5 increase in mAP | Helped |
| Add KFS-stacking temporal fusion (ablation) | MapTR baseline · nuScenes val | mAP: +7.4 | Helped |
| Add KFS-streaming temporal fusion (ablation) | MapTR baseline · nuScenes val | mAP: 7.2 mAP | Helped |
| Add streaming temporal fusion (ablation) | MapTR baseline · nuScenes val | mAP: 3.8 mAP | Helped |
| Add point-wise coordinate embedding (ablation) | Relation embedding module · nuScenes val | mAP: 2.1 | Helped |
| Add point-wise edge-direction embedding (ablation) | Relation embedding module · nuScenes val | mAP: 0.9 | Helped |
| Add instance-wise classification-score embedding (ablation) | Relation embedding module · nuScenes val | mAP: 1.1 | Helped |
| Add instance-wise classification-score and Chamfer-distance embedding (ablation) | Relation embedding module · nuScenes val | mAP: 3.1 | Helped |
| Use Sgiou geometry-aware classification score (ablation) | Geometry-aware alignment · nuScenes val | mAP: -1.0 | Hurt |
| Use Sp2p geometry-aware classification score (ablation) | Geometry-aware alignment · nuScenes val | mAP: 2.8 | Helped |
| Use Sdir geometry-aware classification score (ablation) | Geometry-aware alignment · nuScenes val | mAP: 4.7 | Helped |
Surprising findings
- Using the GIoU-based geometry score alone for the classification target decreased mAP by 1.0 relative to the mask-focal baseline, while point-to-point and direction scores helped.
Limitations
- Stacking variant (KFS-C) has 116.7M parameters versus 40.6M for MapTRv2, and runs slower.
- Evaluated only on three map element categories (lane divider, pedestrian crossing, road boundary) on two datasets.
Experimental setup
Datasets: nuScenes Argoverse2
Benchmarks: nuScenes Argoverse2
Metrics: mAP AP per class (ped crossing, divider, boundary) FPS parameters
Baselines: MapTR MapTRv2 MapVR PivotNet BeMapNet StreamMapNet MGMap SQD-MapNet MapQR HIMap HRMapNet-MapTRv2 VectorMapNet
Builds on: MapTRv2 MapTR BEVFormer SOLOFusion IoU-aware focal loss
Closest papers by meaning
- MapExpert: Online HD Map Construction with Simple and Efficient Sparse Map Element Expert (AAAI 2025 · similarity 0.88)
- Unveiling the Hidden: Online Vectorized HD Map Construction with Clip-Level Token Interaction and Propagation (NeurIPS 2024 · similarity 0.87)
- HIMap: HybrId Representation Learning for End-to-end Vectorized HD Map Construction (CVPR 2024 · similarity 0.87)
- SDTagNet: Leveraging Text-Annotated Navigation Maps for Online HD Map Construction (NeurIPS 2025 · similarity 0.87)
- Uncertainty-Instructed Structure Injection for Generalizable HD Map Construction (CVPR 2025 · similarity 0.87)
- Leveraging SD Map to Augment HD Map-based Trajectory Prediction (CVPR 2025 · 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.