---
configs:
- config_name: default
  data_files:
  - split: full
    path: data/full-*
  - split: curated_10
    path: data/curated_10-*
  - split: thumbnail_3
    path: data/thumbnail_3-*
task_categories:
- image-to-image
pretty_name: Triangle Reconstruction Benchmark
size_categories:
- n<1K
---

# Triangle Reconstruction Benchmark

A small, diverse dataset curated manually from existing datasets and images found online, with item-level rights and provenance information.

The purpose of this dataset is to evaluate the quality of algorithms for approximating images using sets of coloured triangles, such as [Genetic Algorithms](https://www.rogeralsing.com/2008/12/07/genetic-programming-evolution-of-mona-lisa/) or greedy ones such as [Geometrize](https://www.geometrize.co.uk/). 

## Composition

The benchmark set consists of 75 images, all rendered and post-processed to RGB PNG with a 1024-pixel longest edge. Aspect ratio is preserved. Exact source files are retained under `originals/`.

Each image has a stable integer `index` from 1 to 75 for convenient reference (for example, “sample 23”), as well as a content-stable hexadecimal `id`. The same image retains its full-set index in every subset.

There are three sets: 

- `full`: the entire 75-image collection, 
- `curated_10`: a 10-image development subset useful when iterating on new algorithms, and
- `thumbnail_3`: three visually distinct headline images selected for thumbnails and quick comparisons. This includes the Mona Lisa, a target famously used by projects such as [EvoLisa](https://www.rogeralsing.com/2008/12/07/genetic-programming-evolution-of-mona-lisa/).

## Rights and attribution

This dataset contains images from multiple sources. No new or collective license is asserted over the source images; each image remains subject to its original license or rights status. Source, rights, and attribution information is provided at the item level. Canonical images have been resized and/or rasterized for evaluation.

See `THIRD_PARTY_NOTICES.md`, `third_party_licenses/`, and the Parquet metadata columns. In particular, DIV2K is described by its publisher as academic-research-only, Kodak has no first-party license file in the acquired mirror, and logos may remain subject to trademark rights.

## Loading

```python
from datasets import load_dataset

dataset = load_dataset("benchislett/TrianglePaintBench")
```

When working from a local clone, use `load_dataset("path/to/triangle_reconstruction_hf")` instead.

Integrity hashes for the original and canonical image files are in `checksums.sha256`.