Map Query Bank: A New Map Representation for Online Map Generation in Autonomous Driving

Ziming Liu¹, Leichen Wang¹, Ge Yang^1,2, Xinrun Li^1,3, Xingtao Hu¹, Hao Sun¹, Guanyu Gao⁴

¹ AID-OMG Team, Bosch Research, Shanghai, China, ² University of Stuttgart, Stuttgart, Germany, ³ Newcastle University, Newcastle upon Tyne, England, ⁴ Beijing Institute of Technology, Beijing, China

submitted to IEEE Robotics and Automation Letters (RA-L)

---

🚗 Overview

High-definition (HD) maps are essential for autonomous driving, providing the detailed and accurate map information needed for planning and navigation.
However, building and maintaining HD maps is complex, expensive, and time-consuming, making it difficult to scale across diverse environments.

To address this challenge, we propose a Online Map Generation (OMG) framework. Our method leverages low-cost and readily available Standard-definition (SD) maps as strong priors to enable efficient and reliable HD map detection and generation on-the-fly.

---

📦 Our Contributions

✅ Generalizable MQBank

• A new map representation that learns feature embeddings rather than relying on fixed position encoding.

✅ SD Map Prior Query Initialization

• The first approach using SD maps for dynamic map query initialization, breaking the performance ceiling of existing Transformer-based OMG models.

✅ Attention Layer Enhancements

• Incorporates MQBank directly into map decoder attention layers, enabling better local point feature capture.

✅ Extended OpenLaneV2 Dataset

• A manually verified and corrected SD map subset for OpenLaneV2, publicly available to the research community.

---

🔥 Experimental Highlights

• Achieved state-of-the-art performance on OpenLaneV2 benchmark:
  • 40.5% mAP on vehicle lanes
  • 45.7% mAP on pedestrian areas
• New insight into the impact of SD map prior quality on HD map generation.

---

📦 Dataset Release

We release our extended OpenLaneV2 Dataset with high-quality SD map annotations:

🔗 Download Links

Resource	Link
OpenLaneV2 Dataset	GitHub Link
Our Extended SD Map Subset (train)	Google Drive
Our Extended SD Map Subset (val)	Google Drive

📁 Dataset Structure

After downloading the Extended SDMap Subset, the folder structure is as follows:

sdmap_train_modified/
├── 00000/
├── 00001/
├── ...
├── 00699/
└── info_train.json

sdmap_val_modified/
├── 10000/
├── 10001/
├── ...
├── 10149/
└── info_val.json

📄 info_train.json and info_val.json Description

These JSON files provide a validity index for each scene log in the dataset.
Each key corresponds to a scene folder (e.g., 00000, 00001, ...), and the value indicates whether the scene contains a usable and verified SD map.

---

Qualitative analysis of extended sdmap subset

The following images showcase the comparison between the original SDMap and the modified SDMap for four different scenarios.

• Gray dashed lines represent the ground truth HDMap.
• Colored solid lines represent the SDMap.

Scenario 1:

Scenario 2:

Scenario 3:

---

Original vs Modified SDMap (Animated Comparison):

• Left GIF: Original SDMap
• Right GIF: Modified SDMap

---

Using GUI to Edit SDMap

This section explains the step-by-step process of editing an SDMap using the GUI interface. Below are the detailed steps:

---

1. Basic Interface

The GUI provides a user-friendly interface to load and edit SDMaps. Upon starting the application, you will see the following main screen:

---

2. Select log_id

Select the desired log_id corresponding to the specific SDMap you wish to edit.

---

3. Edit the Number of Lanes

Select a road in the map to modify the lane number. The lane number field is editable, allowing you to input the correct number of lanes.

---

4. Delete a Road

To remove a road from the map, select the road and click the Delete Road button.

---

5. Save Edited Lane Numbers

After making changes to the lane numbers or deleting roads, the updated information will be written back to the SDMap.

---

6. Jump to the Next log_id

Once you've finished editing the current log_id, click the Next button to proceed to the next map for editing. This allows for efficient processing of multiple maps.

---

📖 Citation

If you find our paper or dataset helpful, please consider giving us a ⭐ and citing our work:

@misc{liu2025general,
    title        = {Map Query Bank: A New Map Representation for Online Map Generation in Autonomous Driving},
    author       = {Ziming Liu and Leichen Wang and Ge Yang and Xinrun Li and Xingtao Hu and Hao Sun and Guanyu Gao},
    year         = {2025},
    publisher    = {GitHub},
    howpublished = {\url{https://github.com/LaoWangBosch/Map_Query_Bank}},
    note         = {Extended OpenLaneV2 Dataset with SD map annotations available at GitHub}
}