---
license: mit
tags:
- geometry
- fabrication
- symbolic-geometry
- cnc
- sdf
pretty_name: MiGumi Dataset
language:
- en
size_categories:
- n<1K
---

# MiGumi Dataset: Dataset of Millable Kigumi Joints

![assets/teaser.png](assets/teaser.png)

This dataset accompanies the paper **"MiGumi: Making Tightly Coupled Integral Joints Millable" (SIGGRAPH Asia 2025)**  
and provides symbolic, geometric, and mesh representations of traditional Japanese joints as well as their redesigned variants for CNC milling.

---

## Dataset Structure

All the joints are stored in `joints` folder.
Each joint folder (e.g., `CJ_AKT/`) contains:

```markdown
base.json                 # The manually created input Joint Design in MXG language.
info.json                 # High-level metadata for the joint
polyline_files/           # MXG representation for the different variants
vis_files/                # Conversion into a export friendly format (used with blender)
meshes/                   # STL meshes for each variant created using Trimesh & Shapely
```

Variants include:

- `base`: original joint geometry  

- `mill`: Milling Only baseline: Applies morphological opening to each part’s extruded region, producing millable geometry by construction. No attempt is made to preserve surface coupling.

- `odf`: Opening & Diff-Flip: Begins with millable parts obtained via opening, then restores contact by applying the resulting shape differences to paired parts. That is, the volume of each part removed due to the opening operation (Diff) is directly added to the paired part (Flip). Although simple, this heuristic often yields invalid subtractions that break millability.

- `ours`: Our Method which optimizes the joint designs with losses that model tight coupling between joints along with milling.

Generally, ODF results in cleaner designs with smoother milling contours. However, there are two issues:

- It fails for various designs (check out `CJ_SAT`, `CJ_KMKT` etc.).

- It sometimes contains thin tiny slivers in the outline due to issues with the polycurve boolean operations.

In contrast, our approach provides a feasible solution for all the joint designs considered above.

---

## Usage

You can clone the dataset directly:

```bash
git lfs install
git clone https://huggingface.co/datasets/bardofcodes/migumi-dataset
```

Or load it with the Hugging Face Datasets API:

```python
from datasets import load_dataset
ds = load_dataset("bardofcodes/MiGumi-Dataset")
```

(If the dataset is structured by folders, the library will list joint names and file paths.)

---

## Dataset Overview

![assets/full_dataset](assets/full_dataset.png)

---

## License

This dataset is released under the [MIT License](./LICENSE).

---

## Citation

If you use this dataset, please cite:

```markdown
@dataset{ganeshan2025_migumi_dataset,
  author    = {Ganeshan, Aditya and Fleischer, Kurt and Jakob, Wenzel
                and Shamir, Ariel and Ritchie, Daniel 
                and Igarashi, Takeo and Larsson, Maria},
  title     = {MiGumi Dataset: Millable Kigumi Joints Dataset},
  year      = {2025},
  note      = {Available at https://huggingface.co/datasets/bardofcodes/MiGumi-Dataset}
}
```

---

## Related Resources

- **Project page:** [https://bardofcodes.github.io/migumi](https://bardofcodes.github.io/migumi)  
- **Paper:** (to appear in ACM TOG 2025)  
- **Code:** [https://github.com/bardofcodes/migumi](https://github.com/bardofcodes/migumi)