SpaRRTa-Attention: Attention-Analysis Split of the SpaRRTa Benchmark

SpaRRTa-Attention is the interpretability asset for the synthetic SpaRRTa benchmark. Each scene ships with per-object segmentation masks so that a frozen Visual Foundation Model's self-attention can be measured between the objects in the scene (Human / Tree / Truck), the CLS token, the background, and register tokens.

• 📄 Paper: arXiv:2601.11729
• 💻 Code: github.com/gmum/SpaRRTa (see sparrta/analysis/)
• 🧩 Main (synthetic) split: turhancan97/SpaRRTa

Contents

• Total scenes: 300 across 5 environments.
• One folder per environment; one folder per scene, each with the rendered image and its masks:

environment	scenes
bridge	60
city	60
desert	60
forest	60
winter_town	60

<environment>/params_XXXX/
  img_XXXX.jpg
  metadata/
    mask_Human.png
    mask_Tree.png
    mask_Truck.png
    masks_log.csv

Masks are binary PNGs aligned to the image; masks_log.csv records the per-object mask metadata.

Generated on 2026-06-26T09:11:19.002413+00:00.

Use with the SpaRRTa code

Download the dataset, then point the analysis code at it:

huggingface-cli download turhancan97/SpaRRTa-Attention --repo-type dataset --local-dir ./hf_SpaRRTa-Attention
export SPARRTA_ANALYSIS_ROOT=$(pwd)/hf_SpaRRTa-Attention

Then run the attention analysis (see the code repository):

python sparrta/analysis/compute_attention.py environment=winter_town

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}
}