GIQ Benchmark

Dataset Details

GIQ is a benchmark for evaluating geometric reasoning in vision and vision-language models using polyhedra.

The repository currently contains:

• synthetic_images/: synthetic renderings of polyhedra
• wild_images/: real photographs organized by source collection
• 3d_meshes/: 3D mesh assets

The benchmark is designed to support evaluation on tasks such as:

• zero-shot polyhedron classification
• symmetry-related reasoning
• mental rotation
• monocular 3D reconstruction

Uses

This dataset is intended for (but not limited to):

• benchmarking geometric reasoning in computer vision
• evaluating robustness of foundation models on controlled 3D shape data
• comparing synthetic and real-image performance

Dataset Creation

Curation Rationale

The dataset was created to provide a controlled benchmark for geometric reasoning using polyhedra, with both synthetic renderings and real photographs.

Source Data

Data Collection and Processing

Synthetic images were rendered from 3D polyhedron meshes. Real images were collected as photographs of physical polyhedron models.

Who are the source data producers?

The dataset authors created the synthetic portion and curated the benchmark assets.

Personal and Sensitive Information

This dataset is not intended to contain personal or sensitive information.

Citation

If you use this dataset, please cite:

@inproceedings{michalkiewicz2026giq,
  title={GIQ: Benchmarking 3D Geometric Reasoning of Vision Foundation Models with Simulated and Real Polyhedra},
  author={Michalkiewicz, Mateusz and Sokhal, Anekha and Michalkiewicz, Tadeusz and Pawlikowski, Piotr and Baktashmotlagh, Mahsa and Jampani, Varun and Balakrishnan, Guha},
  booktitle={International Conference on Learning Representations (ICLR)},
  year={2026}
}