Research map / Point clouds / Point cloud classification
HIMap: HybrId 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 HIMap, a hybrid point-element representation (HIQuery) with a point-element interactor and consistency loss for end-to-end vectorized HD map construction, reaching 73.7 mAP (ResNet50, 110 epochs) on nuScenes and 77.8 with LiDAR fusion.
Problem
Vectorized HD map construction predicts class and point coordinates of map elements, but point-level representation pipelines lack element-level information and interaction, causing erroneous element shapes and inter-element entanglement.
Approach
A hybrid query (HIQuery) holding point queries and an element query per map element, updated by a point-element interactor that mutually exchanges point-level and element-level features, plus a point-element consistency loss that aligns the two levels via a pseudo element representation and binary cross-entropy on an element similarity matrix.
Key results
- On nuScenes val (ResNet50, 110 epochs, multi-view images), HIMap reaches 73.7 mAP under the easy setting, +5.0 over MapTRv2, and 51.6 mAP under hard setting, +4.5 over BeMapNet.
- With multi-modality input (images + LiDAR), HIMap achieves 77.8 mAP on nuScenes at 110 epochs, 8.3 mAP above MapTR-based prior results.
- On Argoverse2 val (24 epochs), HIMap reaches 72.7 mAP easy and 50.0 mAP hard, exceeding MapTRv2 by 2.2 and 3.5 mAP respectively.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| replace point-level representation with hybrid representation and element-level masked attention (ablation) | HIMap · nuScenes val, easy | mAP: 68.5 -> 70.6 | Helped |
| add point-element interaction (ablation) | HIMap with hybrid representation · nuScenes val, easy | mAP: 70.6 -> 73.1 | Helped |
| add point-element consistency constraint (ablation) | HIMap with point-element interaction · nuScenes val, easy | mAP: 73.1 -> 73.7 | Helped |
| share position embeddings between point and element feature extractors (ablation) | HIMap point-element interactor · nuScenes val, easy | mAP: 70.6 -> 71.6 | Helped |
| update only the point query with integrated information (ablation) | HIMap point-element interactor with shared positions · nuScenes val, easy | mAP: 71.6 -> 72.1 | Helped |
| update only the element query with integrated information (ablation) | HIMap point-element interactor with shared positions · nuScenes val, easy | mAP: 71.6 -> 72.8 | Helped |
| update both point and element queries with integrated information (ablation) | HIMap point-element interactor with shared positions · nuScenes val, easy | mAP: 71.6 -> 73.1 | Helped |
| set point-element consistency loss weight to 0.0 instead of 2.0 (ablation) | HIMap · nuScenes val, easy | mAP: 73.7 -> 73.1 | Hurt |
| set point-element consistency loss weight to 1.0 instead of 2.0 (ablation) | HIMap · nuScenes val, easy | mAP: 73.7 -> 73.5 | Hurt |
| increase point-element consistency loss weight to 3.0 (ablation) | HIMap · nuScenes val, easy | mAP: 73.7 -> 72.5 | Hurt |
| use HIMap instead of BeMapNet | ResNet50 vectorized map construction · nuScenes val, hard | mAP: 47.1 -> 51.6 | Helped |
| use HIMap instead of MapTRv2 | ResNet50 vectorized map construction · nuScenes val, easy | mAP: 68.7 -> 73.7 | Helped |
Surprising findings
- On Argoverse2 the divider class AP is slightly lower than MapTRv2 under the easy threshold setting, though higher under the strict 0.2m threshold.
Limitations
- Focuses on accuracy; model acceleration is left for future work.
- Only constructs 2D HD maps, ignoring height变化 of the road.
- Does not model temporal consistency of HD maps across timestamps.
Experimental setup
Datasets: nuScenes Argoverse2
Benchmarks: nuScenes Argoverse2
Metrics: mAP (chamfer distance based) AP per class
Baselines: HDMapNet VectorMapNet MapTR MapVR BeMapNet MapTRv2 PivotNet
Builds on: MapTR MapTRv2 Deformable DETR Masked attention DAB-DETR
Closest papers by meaning
- InteractionMap: Improving Online Vectorized HDMap Construction with Interaction (CVPR 2025 · similarity 0.87)
- Compact HD Map Construction via Douglas-Peucker Point Transformer (AAAI 2024 · similarity 0.86)
- MGMapNet: Multi-Granularity Representation Learning for End-to-End Vectorized HD Map Construction (ICLR 2025 · similarity 0.86)
- MGMap: Mask-Guided Learning for Online Vectorized HD Map Construction (CVPR 2024 · similarity 0.85)
- Unveiling the Hidden: Online Vectorized HD Map Construction with Clip-Level Token Interaction and Propagation (NeurIPS 2024 · similarity 0.85)
- Uncertainty-Instructed Structure Injection for Generalizable HD Map Construction (CVPR 2025 · 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.