---
license: apache-2.0
tags:
- cuneiform
- msii
- multi-scale-integral-invariant
- curvature
- 3d
- image
pretty_name: Cuneiform Photos⇔MSII
---

<div align="center">

<h1 align="center">
Cuneiform Photos⇔MSII
</h1>

<img src="./assets/PhotoToMSII.png" width="300"/>

</div>

This dataset contains paired images of photorealistic cuneiform tablet renders and their corresponding MSII (Multi Scale Integral Invariant) curvature visualizations. Both are rendered in Blender at 4096px max axis.

## Background & Motivation

Cuneiform tablets contain impressions made in clay thousands of years ago. These subtle surface variations are often difficult to see in regular photographs, especially under poor lighting conditions. MSII (Multi Scale Integral Invariant) filtering is a curvature visualization technique that highlights these impressions by computing surface curvature at multiple scales, making cuneiform characters clearly visible regardless of lighting.

However, getting the MSII visualization of a tablet requires a 3D scan and lots of computation. To reduce this barrier and increase the availability of easy-to-read images, I'd like to train a diffusion model to predict the MSII visualization directly from photographs.

To do that, I've created this high quality dataset intended for training.

## Dataset Format

| Column | Description | Example |
|--------|-------------|---------|
| hs_number | The Heidelberg Sample identifier | "HS_0044" |
| variation | The variant index for this tablet | 1-8 |
| tablet_url | A link to the tablet data | "https://hilprecht.mpiwg-berlin.mpg.de/object3d/XXXXX" |
| cdli_archive | A link to the tablet information in the Cuneiform Digital Library, if applicable | "https://cdli.ucla.edu/search/archival_view.php?ObjectID=PXXXXXX" |
| photo | A synthetic photograph of the tablet | <img src="./assets/example_photo.png"/> |
| msii | A matching MSII visualization of the tablet | <img src="./assets/example_msii.png"/> |

## Source Data

This project uses the HeiCuBeDa (Heidelberg Cuneiform Benchmark Dataset), a professional research dataset of 1,747 high-resolution 3D scans.

## Image Diversity

To get the most out of the 1.7K tablet scans from HeiCuBeDa, we generate several varied images for each tablet.

Each variant gets an independent random selection of:

| What Varies | Description | Range |
|-------------|-------------|-------|
| Faces | Which of the six faces are shown in the image | Variant 1 always shows 6 faces. Variants 2+ have a 20% chance of all 6, otherwise uses per-face probs (front 100%, back 75%, top/bottom 45%, left/right 30%) |
| Rotation | Rotation of the tablet | ±5° Euler XYZ |
| Focal length | The focal length of the camera | 80–150 mm |
| Lighting | The position, color, and intensity of the lights in the photo | ±30% Energy. ±2% Warmth. Random offset along the perpendicular plane |
| Background | The image behind the tablet in the photo | 70% use no background. 30% select from fabric/grunge/stone (weighted). Rotation/scale/offset are randomized. Perlin noise is used to vary brightness throughout. |

### Known Limitations

- Renders may not capture all real-world photo degradation
- MSII visualization quality depends on PLY mesh resolution and artifacting
- Generalization to real photos vs. renders needs validation

# Citations & Acknowledgments

We thank the digital humanities and archaeology communities for their foundational work in cuneiform digitization and analysis.

- Mara, H. (2019). HeiCuBeDa Hilprecht - Heidelberg Cuneiform Benchmark Dataset for the Hilprecht Collection (Version V2) dataset. https://doi.org/doi:10.11588/DATA/IE8CCN
- Bogacz, Bartosz & Mara, Hubert. (2018). Feature Descriptors for Spotting 3D Characters on Triangular Meshes. https://doi.org/doi:10.1109/ICFHR-2018.2018.00070
- Bayer, V. and Mara, H. (2019). GigaMesh Software Framework Tutorial 6: Screenshot Rendering. https://doi.org/10.11588/heidok.00026537
- Blender Foundation. (2025). Blender (Version 4.4). https://www.blender.org