leggedrobotics/defm-bench-graspnet

DeFM benchmark: GraspNet-1Billion

Depth semantic segmentation of tabletop objects: 2000 train / 600 test frames, 40 object classes (background is ignored).

This is the version used in the benchmarks of DeFM: Learning Foundation Representations from Depth for Robotics (2,000 train, 600 test). The original data is from GraspNet-1Billion; 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 graspnet

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 / 1000, 0 = invalid
label	image	8-bit PNG, class ids 0..39, 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{fang2020graspnet, title={GraspNet-1Billion: A Large-Scale Benchmark for General Object Grasping}, author={Fang, Hao-Shu and Wang, Chenxi and Gou, Minghao and Lu, Cewu}, booktitle={CVPR}, year={2020}}