OMA-MAT
[ICLR 2026] Online Navigation Refinement: Achieving Lane-Level Guidance by Associating Standard-Definition and Online Perception Maps
Online Map Association Benchmark and Framework
Connecting online mapping with hybrid navigation to enable interpretable autonomous driving.
Key insights:
- We introduce Online Map Association (OMA), the first benchmark for hybrid navigation-oriented online map association.
- We introduce Association P-R, a metric for map association that considers the accuracy and precision of topological alignment.
- We propose a Map Association Transformer (MAT), which utilizes path-aware attention and spatial attention mechanisms to enable understanding of geometric and topological correspondences.
News
- 2026/02/27: 🚀 Dataset and checkpoint are now available on HuggingFace and ModelScope.
- 2026/01/26: 🎉 Congratulations, OMA-MAT is accepted by ICLR 2026. We will open the dataset and checkpoint as soon as possible.
- 2025/07/15: First commit.
Quick Start
Prepare Dataset
# Download the OMA dataset to the data/oma directory using the Huggingface CLI:
huggingface-cli download wanjiaxu/OMA --repo-type dataset --local-dir data/oma
Training
# By script (Recommended)
# -p is default set as python and can be ignored
sh scripts/train.sh -p python -d oma -c oma-mt-v1m1-l -n oma-mt-v1m1-l
# Direct
export PYTHONPATH=./
python tools/train.py --config-file configs/oma/oma-mt-v1m1-l.py --options save_path=exp/oma/oma-mt-v1m1-l
Test
# By script (Based on experiment folder created by training script)
# -p is default set as python and can be ignored
# -w is default set as model_best and can be ignored
sh scripts/test.sh -p python -d oma -n oma-mt-v1m1-l -w model_best
# Direct
export PYTHONPATH=./
python tools/test.py --config-file configs/oma/oma-mt-v1m1-l.py --options save_path=exp/oma/oma-mt-v1m1-l weight=exp/oma/oma-mt-v1m1-l/model/model_best.pth
To use the pretrained checkpoint from HuggingFace directly:
# Download the pretrained checkpoint
huggingface-cli download wanjiaxu/MAT --local-dir checkpoints/MAT
# Run evaluation with the downloaded checkpoint
export PYTHONPATH=./
python tools/test.py --config-file configs/oma/oma-mt-v1m1-l.py --options save_path=exp/oma/oma-mt-v1m1-l weight=checkpoints/MAT/model_best.pth
Evaluate with Association P-R Metric
After testing, the model outputs prediction JSON files. Run the Association P-R metric evaluation using the scripts in metrics/:
cd metrics
# Step 1: Compute per-sample TP/FP/FN statistics
python metrics.py \
--file_dir ../exp/oma/oma-mt-v1m1-l/result \
--output_dir ../exp/oma/oma-mt-v1m1-l/metric_result \
--gt_dir ../data/oma/val \
--distance_threshold 1.0
# Step 2: Aggregate results and print P / R / F1
python read_and_recal_metric.py \
--file_dir ../exp/oma/oma-mt-v1m1-l/metric_result
Model and Dataset
| Resource | HuggingFace | ModelScope |
|---|---|---|
| Dataset (OMA) | 🤗 wanjiaxu/OMA | WallelWan/OMA |
| Checkpoint (MAT) | 🤗 wanjiaxu/MAT | WallelWan/MAT |
Licence
This project is released under MIT licence.
Acknowledgment
This project is mainly based on the following projects:
The Readme is inspired by DeepEyes.
TODO
- Improved documentation and tutorials
- Open resource the post-process code.
- Open resource the eval metric code.
- Open resource the dataset in huggingface.
- Open resource the checkpoint in huggingface.
Citation
@article{wan2025online,
title={Online Navigation Refinement: Achieving Lane-Level Guidance by Associating Standard-Definition and Online Perception Maps},
author={Wan, Jiaxu and Wang, Xu and Xie, Mengwei and Chang, Xinyuan and Liu, Xinran and Pan, Zheng and Xu, Mu and Zhang, Hong and Yuan, Ding and Yang, Yifan},
journal={arXiv preprint arXiv:2507.07487},
year={2025}
}
}