mm305/giq

# 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:

```bibtex
@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}
}