artest08/TopoMaskV3-benchmark

TopoMaskV3 Benchmark: Geographically Disjoint and Long-Range (±100 m) Splits

Additional annotation and pre-built pkl files for the benchmark introduced in TopoMaskV3: 3D Mask Head with Dense Offset and Height Predictions for Road Topology Understanding, accepted to CVPR 2026 Workshops (AUTOPILOT 2026).

• Paper: arXiv:2603.01558
• Project page: artest08.github.io/TopoMaskV3.github.io
• Code and split definitions: OpenLane-V2 fork, different_splits branch

These files extend OpenLane-V2 Subset-A (built on Argoverse 2). They contain annotations only. Camera images and other sensor data are not included: download the original OpenLane-V2 Subset-A dataset separately.

Benchmark

Standard OpenLane-V2 evaluation draws train and evaluation samples from the same geographic locations. This benchmark adds two stricter protocols:

1. Geographically disjoint splits: train and evaluation sets share no geographic overlap, so models must generalize to unseen roads.
2. Long-range (±100 m) evaluation: extends the standard ±50 m perception range.

Split	Geographic overlap	City overlap	Evaluation focus	±100 m variant
Original	Yes	Yes	In-distribution performance	Subset-A-100
Near	No	Yes	Within-city generalization	Subset-A-near-100
FarA / FarB / FarC	No	No	Geographic generalization	Subset-A-farA-100 / -farB-100 / -farC-100

Files

File	Contents	Size	md5
openlanev2_sA_test_anno.zip	Ground-truth annotations for the test split (±50 m), creates test_gt/	1.67 GB	3beaeef0e41590fedebfd9717ca5e590
openlanev2_sA_anno_100.zip	Re-annotated info/ files for the ±100 m splits, creates full_json_100/	18.8 GB	e9ce6d9f1469a8f31eecff544904b09c
openlanev2_sA_pkl_files.zip	Pre-built pkl files for all splits (data_dict_subset_A_*.pkl)	20.7 GB	388d3646db826b769a7bb91275981252

openlanev2_sA_pkl_files.zip is still being uploaded. Until it appears in this repo, download it from the Google Drive folder.

All three archives share the same internal root, datasets/OpenLane-V2/.

If the pre-built pkl files are enough for your use case, download only openlanev2_sA_pkl_files.zip and skip the Data Setup steps in the fork README.

Download

With the Hugging Face CLI (pip install -U huggingface_hub):

# Everything
hf download artest08/TopoMaskV3-benchmark --repo-type dataset --local-dir ./topomaskv3_benchmark

# A single file
hf download artest08/TopoMaskV3-benchmark openlanev2_sA_test_anno.zip --repo-type dataset --local-dir ./topomaskv3_benchmark

Or with a direct link:

wget https://huggingface.co/datasets/artest08/TopoMaskV3-benchmark/resolve/main/openlanev2_sA_test_anno.zip

Check the download against the md5 in the table above:

md5sum openlanev2_sA_test_anno.zip

Setup

Extract each zip, then move its contents into your OpenLane-V2 dataset root:

EXTRACT_DIR=/path/to/extracted

mv "$EXTRACT_DIR/datasets/OpenLane-V2/test_gt" /path/to/OpenLane-V2/
mv "$EXTRACT_DIR/datasets/OpenLane-V2/full_json_100" /path/to/OpenLane-V2/
mv "$EXTRACT_DIR"/datasets/OpenLane-V2/data_dict_subset_A*.pkl /path/to/OpenLane-V2/

Then follow the Data Setup steps in the fork README to link the test annotations and build the merged/ and merged_100/ folders.

License

These annotations are derived from OpenLane-V2 and are distributed under the same CC BY-NC-SA 4.0 license. Before using them, you must agree to the terms of use of Argoverse 2, the source of OpenLane-V2 Subset-A.

Citation

@article{kalfaoglu_topomaskv3_2026,
  title={TopoMaskV3: 3D Mask Head with Dense Offset and Height Predictions for Road Topology Understanding},
  journal={arXiv preprint arXiv:2603.01558},
  author={Kalfaoglu, Muhammet Esat and Ozturk, Halil Ibrahim and Kilinc, Ozsel and Temizel, Alptekin},
  year={2026}
}

@inproceedings{wang2023openlanev2,
  title={OpenLane-V2: A Topology Reasoning Benchmark for Unified 3D HD Mapping},
  author={Wang, Huijie and Li, Tianyu and Li, Yang and Chen, Li and Sima, Chonghao and Liu, Zhenbo and Wang, Bangjun and Jia, Peijin and Wang, Yuting and Jiang, Shengyin and Wen, Feng and Xu, Hang and Luo, Ping and Yan, Junchi and Zhang, Wei and Li, Hongyang},
  booktitle={NeurIPS},
  year={2023}
}