---
license: mit
tags:
  - autonomous-driving
  - hd-map
  - vectorized-map
  - nuscenes
  - mmdetection3d
datasets:
  - nuscenes
---

# MapGR: MapQR + MapGR checkpoint (nuScenes, ResNet-50, 24 epochs)

This repository hosts the released weights for **Learning Global Representation from Queries for Vectorized HD Map Construction** (ICML 2026, [arXiv:2510.06969](https://arxiv.org/abs/2510.06969), [project page](https://xr-lee.github.io/mapgr-demo/)).

Training and evaluation code, installation and data preparation are in the GitHub repository [XR-Lee/MapGR](https://github.com/XR-Lee/MapGR) (`mapqr/` sub-project).

| File | Description |
|---|---|
| `mapqr_mapgr_nusc_r50_24ep.pth` | Model weights: `state_dict` plus minimal `meta`, no optimizer state. |
| `mapqr_mapgr_nusc_r50_24ep_config.py` | Fully resolved config, for reference. Train and evaluate with `mapqr/projects/configs/mapgr/mapqr_mapgr_nusc_r50_24ep.py` from the code repository. |

## Results (nuScenes val)

| | mAP<sub>1</sub> | mAP<sub>2</sub> | AP<sub>div</sub> / AP<sub>ped</sub> / AP<sub>bou</sub> (mAP<sub>2</sub> thresholds) |
|---|---|---|---|
| Paper (MapQR + MapGR, 24 ep) | 45.3 | 68.1 | 69.8 / 65.8 / 68.8 |
| This checkpoint | 44.87 | 67.96 | 69.35 / 64.95 / 69.57 |

mAP<sub>2</sub> averages Chamfer AP over the {0.5, 1.0, 1.5} m thresholds; mAP<sub>1</sub> averages over {0.2, 0.5, 1.0} m.

This checkpoint is not the run reported in the paper.

- It was trained with the research code the GitHub repository was cleaned up from, using the same settings as the released config.
- Training: 8 GPUs × 4 samples, fp16.
- The numbers above come from evaluating this exact file with the released `tools/test.py`.
- Other 24-epoch runs of closely related configurations (different seeds, loss weights and query counts) reached 67.2–68.0 mAP<sub>2</sub>.

```shell
# inside mapqr/ of the code repository
wget -P ckpts https://huggingface.co/Lee1413/MapGR/resolve/main/mapqr_mapgr_nusc_r50_24ep.pth
./tools/dist_test_map.sh projects/configs/mapgr/mapqr_mapgr_nusc_r50_24ep.py ckpts/mapqr_mapgr_nusc_r50_24ep.pth 1
```

SHA-256 of `mapqr_mapgr_nusc_r50_24ep.pth`: `d4f868312234806b5c0a97ea2c7edf71173de3fac1d510742a99317e1c6f6916`

## Training logs

mmdet3d logs of the two 24-epoch nuScenes runs. User names, host names and internal path prefixes are replaced with `<user>`, `<host>` and `<workspace>`. Nothing else is changed.

| Run | Files | Setup | mAP<sub>2</sub> at epoch 12 / 24 |
|---|---|---|---|
| MapQR + MapGR (this checkpoint) | [`20260329_014512.log`](logs/mapqr_mapgr_nusc_r50_24ep/20260329_014512.log), [`.log.json`](logs/mapqr_mapgr_nusc_r50_24ep/20260329_014512.log.json) | 8 GPUs × 4 samples, fp16, PyTorch 2.0.1 | 59.90 / 67.96 |
| MapTRv2 + MapGR (the paper's run; weights not released) | [`train.273710.log`](logs/maptrv2_mapgr_nusc_r50_24ep/train.273710.log) | 4 GPUs × 4 samples, fp16, PyTorch 1.9.1 | 56.14 / 64.98 |

## Citation

```bibtex
@inproceedings{qiu2026mapgr,
  title     = {Learning Global Representation from Queries for Vectorized {HD} Map Construction},
  author    = {Qiu, Shoumeng and Li, Xinrun and Long, Yang and Xue, Xiangyang and Ojha, Varun and Pu, Jian},
  booktitle = {International Conference on Machine Learning (ICML)},
  year      = {2026}
}
```