SpaRRTa-Lego: Real-World Split of the SpaRRTa Spatial-Relation Benchmark

SpaRRTa-Lego is the real-world counterpart of the synthetic SpaRRTa benchmark. Scenes are photographed with toy minifigures and everyday objects, and used for sim-to-real evaluation of the spatial-relation capabilities of Visual Foundation Models.

• 📄 Paper: arXiv:2601.11729
• 💻 Code: github.com/gmum/SpaRRTa
• 🧩 Synthetic split: turhancan97/SpaRRTa

The task

A 4-way classification problem — Front / Back / Left / Right — labelling where a target object lies relative to a reference object from the viewpoint of a human (lego) figure in the scene (the allocentric setting). Each image's class is given by its folder.

Classes & size

• Total images: 1,060 (single train split)
• One folder per class:

class	images
front	263
back	261
left	273
right	263

Generated on 2026-06-25T14:55:20.537953+00:00.

Loading

from datasets import load_dataset

ds = load_dataset("turhancan97/SpaRRTa-Lego", split="train")
print(ds.features["label"])   # ClassLabel(names=['back', 'front', 'left', 'right'])
ds[0]["image"]                # PIL.Image (decoded automatically)

Use with the SpaRRTa code

Download the dataset, then point the code's lego dataset at the train/ folder:

huggingface-cli download turhancan97/SpaRRTa-Lego --repo-type dataset --local-dir ./hf_SpaRRTa-Lego

The download reproduces hf_SpaRRTa-Lego/train/{front,back,left,right}/*.jpg, which is exactly the layout expected by LegoRelativePosition / dataset=lego_position:

export SPARRTA_LEGO_ROOT=$(pwd)/hf_SpaRRTa-Lego/train
python train.py \
  backbone=dino_b16 \
  dataset=lego_position \
  probe=classifier probe._target_=sparrta.models.probes.EfficientProbing

See the code repository for split modes (random, time_series, clip_block_random, clip_block_time_series) and full instructions.

License

Released under the MIT License.

Citation

@misc{kargin2026sparrta,
  title={SpaRRTa: A Synthetic Benchmark for Evaluating Spatial Intelligence in Visual Foundation Models},
  author={Turhan Can Kargin and Wojciech Jasiński and Adam Pardyl and Bartosz Zieliński and Marcin Przewięźlikowski},
  year={2026},
  eprint={2601.11729},
  archivePrefix={arXiv},
  primaryClass={cs.CV},
  url={https://arxiv.org/abs/2601.11729}
}