Multi-data Scene Flow Challenge

One method, five LiDAR datasets: Argoverse 2, nuScenes, Waymo, TruckScenes and Aeva.

Start with OpenSceneFlow, KTH RPL's open-source codebase: data processing, training, evaluation and pre-trained models for all five datasets, and one command to write the submission zip.

GitHub Pre-trained models Challenge data

Overview

Predict scene flow for every test scene with one method and one configuration: a single checkpoint for learned models, or the same hyper-parameters for optimization-based methods. Scene IDs are anonymized, so a method is never told which dataset a scene comes from.

• Ranges: 0–35 m for every entry; 35–70 m is optional (skip it if your model crops the point cloud, e.g. at 51.2 m).
• Tracks: Supervised, or Unsupervised: no human labels of real-world data (self-supervision, simulated labels and foundation models are fine).
• Training data: each entry is labelled single-dataset or multi-dataset.

1. Download the data

Download challenge_public from Hugging Face (about 150 GB):

huggingface-cli download kin-zhang/multidata-sf-challenge --repo-type dataset --local-dir ./challenge_data

• index_eval.pkl: list of (scene_id, timestamp) tuples, one per frame to predict.
• {scene_id}.h5: one HDF5 file per scene with all timestamps (for multi-frame context).

2. Generate predictions

With OpenSceneFlow, any supported model writes the submission directly:

python eval.py checkpoint=/path/to/model.ckpt data_mode=test \
  dataset_path=/path/to/challenge_public leaderboard_version=3 save_res=True

With your own implementation, predict the per-point 3D flow for every (scene_id, timestamp) in index_eval.pkl, in ego-motion-removed (relative) form:

# final_absolute_flow = pred_flow + pose_flow  -> submit pred_flow, NOT the absolute flow
save_pred_flow = (final_flow - pose_flow).cpu().numpy()   # shape (N, 3)

Save each frame as an Apache Feather file with columns flow_tx_m, flow_ty_m, flow_tz_m. The file name is the exact integer timestamp from index_eval.pkl, and the number of rows equals the number of points in lidar (ground points included; they are masked on the server).

import pandas as pd
from pathlib import Path

out_dir = Path("submission") / scene_id
out_dir.mkdir(parents=True, exist_ok=True)
pd.DataFrame({"flow_tx_m": flow[:, 0], "flow_ty_m": flow[:, 1], "flow_tz_m": flow[:, 2]}) \
  .to_feather(out_dir / f"{timestamp}.feather")

Zip the contents of submission/ so the scene folders are at the top level of the zip:

submission.zip
└── {scene_id}/
    ├── {timestamp_0}.feather
    └── ...

cd submission && zip -r -0 ../submission.zip .

3. Upload

Register your team, confirm your e-mail and wait for an organizer to approve the account. Then upload on the Submit page in your browser, or from a terminal with an API key from your profile. The terminal uploader needs only bash and curl and resumes interrupted uploads:

curl -O https://sceneflow.kth-rpl.se/download/sceneflow_submit.sh
export SCENEFLOW_API_KEY=YOUR_API_KEY
bash sceneflow_submit.sh submission.zip --method-name "MyMethod" --training-data multi \
  --project-url "https://github.com/user/repo" --notes "training data, backbone, ..." --wait
# optional: --long-range (also evaluate 35-70 m), --unsupervised, --private; see --help

The uploader checks the zip layout locally before sending anything. After the upload the server validates the file names against the evaluation index and evaluates the submission; results usually appear within 10 minutes. Track the status under My submissions.

Evaluation

Submissions are scored with Dynamic Bucket-Normalized EPE (Khatri et al., ECCV 2024), the metric of the previous AV2 scene flow challenges. It is class-aware and speed-normalized, so slow pedestrians and fast vehicles count fairly. The classic three-way EPE (FD / FS / BS) is reported as well.

Distances are Chebyshev distances in XY (max(|x|, |y|)). For each range and dataset we report mean Static, mean Dynamic, Threeway EPE and per-class EPE for BACKGROUND, CAR, OTHER_VEHICLES, PEDESTRIAN and WHEELED_VRU.

Ranking (lower is better): Overview 0–70 m (default; mean of both ranges), 0–35 m (every entry) and 35–70 m, all by mean Dynamic averaged over the five datasets. Evaluating 35–70 m is your choice when you submit (--long-range) and can be changed later; entries without it are listed below the Overview and 35–70 m rankings instead of being compared, and their 35–70 m numbers stay private.

Ground-truth flow is derived from 3D box tracks with rigid-body assignment. Ground points are excluded, and only frames with at least 10,000 non-ground points and valid flow labels are evaluated. The mapping from anonymous scene IDs to datasets is kept private on the server.

Rules

• Use one configuration for all scenes of all datasets: a single checkpoint for learned models, or the same hyper-parameters for optimization-based methods. No per-dataset models or tuning.
• Do not train on any validation set of any dataset.
• Test-time (per-scene) optimization is allowed but must be declared in the submission notes.
• At most 10 submissions per team per 24 hours, one evaluated at a time. Use one account per team.

Citing

If you take part in the challenge or use its results, please cite the OpenSceneFlow works and the evaluation metric:

@inproceedings{zhang2024deflow,
  author={Zhang, Qingwen and Yang, Yi and Fang, Heng and Geng, Ruoyu and Jensfelt, Patric},
  booktitle={2024 IEEE International Conference on Robotics and Automation (ICRA)},
  title={{DeFlow}: Decoder of Scene Flow Network in Autonomous Driving},
  year={2024},
  pages={2105-2111},
  doi={10.1109/ICRA57147.2024.10610278}
}
@inproceedings{zhang2024seflow,
  author={Zhang, Qingwen and Yang, Yi and Li, Peizheng and Andersson, Olov and Jensfelt, Patric},
  title={{SeFlow}: A Self-Supervised Scene Flow Method in Autonomous Driving},
  booktitle={European Conference on Computer Vision (ECCV)},
  year={2024},
  pages={353–369},
  organization={Springer},
  doi={10.1007/978-3-031-73232-4_20},
}
@inproceedings{zhang2025deltaflow,
  title={{DeltaFlow}: An Efficient Multi-frame Scene Flow Estimation Method},
  author={Zhang, Qingwen and Zhu, Xiaomeng and Zhang, Yushan and Cai, Yixi and Andersson, Olov and Jensfelt, Patric},
  booktitle={The Thirty-ninth Annual Conference on Neural Information Processing Systems},
  year={2025},
  url={https://openreview.net/forum?id=T9qNDtvAJX}
}
@article{zhang2025himo,
  title={{HiMo}: High-Speed Objects Motion Compensation in Point Cloud},
  author={Zhang, Qingwen and Khoche, Ajinkya and Yang, Yi and Ling, Li and Mansouri, Sina Sharif and Andersson, Olov and Jensfelt, Patric},
  journal={IEEE Transactions on Robotics},
  year={2025},
  volume={41},
  pages={5896-5911},
  doi={10.1109/TRO.2025.3619042}
}
@inproceedings{zhang2026teflow,
  title = {{TeFlow}: Enabling Multi-frame Supervision for Self-Supervised Feed-forward Scene Flow Estimation},
  author = {Zhang, Qingwen and Jiang, Chenhan and Zhu, Xiaomeng and Miao, Yunqi and Zhang, Yushan and Andersson, Olov and Jensfelt, Patric},
  year = {2026},
  booktitle = {Proceedings of the IEEE/CVF conference on computer vision and pattern recognition},
  pages = {3667-3676},
}
@article{zhang2026synflow,
  title={{SynFlow}: Scaling Up LiDAR Scene Flow Estimation with Synthetic Data},
  author={Zhang, Qingwen and Zhu, Xiaomeng and Jiang, Chenhan and Jensfelt, Patric},
  journal={arXiv preprint arXiv:2604.09411},
  year={2026}
}
@inproceedings{khoche2025ssf,
  title={{SSF}: Sparse Long-Range Scene Flow for Autonomous Driving},
  author={Khoche, Ajinkya and Zhang, Qingwen and Sanchez, Laura Pereira and Asefaw, Aron and Mansouri, Sina Sharif and Jensfelt, Patric},
  booktitle={2025 IEEE International Conference on Robotics and Automation (ICRA)},
  year={2025},
  pages={6394-6400},
  doi={10.1109/ICRA55743.2025.11128770}
}
@inproceedings{khatri2024trackflow,
  author    = {Khatri, Ishan and Vedder, Kyle and Peri, Neehar and Ramanan, Deva and Hays, James},
  title     = {I Can't Believe It's Not Scene Flow!},
  booktitle = {European Conference on Computer Vision (ECCV)},
  year      = {2024}
}

@article{li2025uniflow,
  title={{UniFlow}: Towards Zero-Shot LiDAR Scene Flow for Autonomous Vehicles via Cross-Domain Generalization},
  author={Li, Siyi and Zhang, Qingwen and Khatri, Ishan and Vedder, Kyle and Ramanan, Deva and Peri, Neehar},
  journal={arXiv preprint arXiv:2511.18254},
  year={2025}
}
@article{khoche2026dogflow,
  author={Khoche, Ajinkya and Zhang, Qingwen and Cai, Yixi and Mansouri, Sina Sharif and Jensfelt, Patric},
  journal = {IEEE Robotics and Automation Letters},
  title = {{DoGFlow}: Self-Supervised LiDAR Scene Flow via Cross-Modal Doppler Guidance},
  year = {2026},
  volume = {11},
  number = {3},
  pages = {3836-3843},
  doi = {10.1109/LRA.2026.3662592},
}
@article{kim2025flow4d,
  author={Kim, Jaeyeul and Woo, Jungwan and Shin, Ukcheol and Oh, Jean and Im, Sunghoon},
  journal={IEEE Robotics and Automation Letters},
  title={Flow4D: Leveraging 4D Voxel Network for LiDAR Scene Flow Estimation},
  year={2025},
  volume={10},
  number={4},
  pages={3462-3469},
  doi={10.1109/LRA.2025.3542327}
}
@inproceedings{lin2025voteflow,
  title={VoteFlow: Enforcing Local Rigidity in Self-Supervised Scene Flow},
  author={Lin, Yancong and Wang, Shiming and Nan, Liangliang and Kooij, Julian and Caesar, Holger},
  booktitle={CVPR},
  year={2025},
}