leggedrobotics/defm-bench-tartanground

DeFM benchmark: TartanGround

Depth semantic segmentation in simulated outdoor environments: 800 train / 400 test frames, 6 classes. Depth is clipped at 100 m (sky).

This is the version used in the benchmarks of DeFM: Learning Foundation Representations from Depth for Robotics (800 train, 400 test). The original data is from TartanGround; please cite it (below) together with DeFM.

Use

git clone https://github.com/leggedrobotics/defm && cd defm && pip install -e ".[benchmarks]"
python -m benchmarks.segmentation --encoder defm_vit_s14 --dataset tartanground

The benchmark downloads this dataset automatically. Class names and the depth scale are in meta.json.

Format

data/<split>-XXXXX-of-XXXXX.parquet:

column	type	content
id	string	frame id
depth	image	16-bit PNG, depth in meters = value / 500, 0 = invalid
label	image	8-bit PNG, class ids 0..5, 255 = ignore

Citation

@misc{patel2026defm,
  title={DeFM: Learning Foundation Representations from Depth for Robotics},
  author={Manthan Patel and Jonas Frey and Mayank Mittal and Fan Yang and Alexander Hansson and Amir Bar and Cesar Cadena and Marco Hutter},
  year={2026}, eprint={2601.18923}, archivePrefix={arXiv}, primaryClass={cs.RO}, url={https://arxiv.org/abs/2601.18923}
}

@inproceedings{patel2025tartanground, title={TartanGround: A Large-Scale Dataset for Ground Robot Perception and Navigation}, author={Patel, Manthan and Yang, Fan and Qiu, Yuheng and others}, booktitle={IROS}, year={2025}}