DONUT (Dataset Of MaNifold strUcTures)

This repository contains a dataset of 3D samples made of watertight meshes and corresponding point clouds. Each sample is composed of one or several watertight mesh components and one 8192-point cloud representation.

The dataset contains 29,517 samples in total (23,579 train, 5,938 test), stored in WebDataset format.

Overview

The figure below shows a few samples from the dataset together with their labels.

Contents

.
├── train/              # train-00000.tar ... train-00039.tar  (~1 GB each)
├── test/               # test-00000.tar  ... test-00009.tar
├── metadata.csv        # one row per sample (same fields as the .json entries)
├── splits/             # id lists for the full, reduced and cross-validation splits
└── tools/build_webdataset.py

train/ and test/ follow the full split in splits/full/. Samples are shuffled (seed 0) before sharding, so shards can be read sequentially.

Sample Format

Each sample is identified by its id (the WebDataset __key__) and contains four files:

File	Content	Shape / dtype
{id}.vertices.npy	mesh vertices (all components)	(V, 3) float32
{id}.faces.npy	mesh triangles, indices into vertices	(F, 3) int32
{id}.pcd.npy	point cloud sampled on the mesh	(8192, 3) float32
{id}.json	id, genus, components, sample_code	—

A sample may contain one or several watertight connected mesh components.

Loading

With webdataset (streaming, recommended for training):

import webdataset as wds

url = "https://huggingface.co/datasets/LouisM2001/donut/resolve/main/train/train-{00000..00039}.tar"
ds = wds.WebDataset(url, shardshuffle=True).shuffle(1000).decode()
for s in ds:
    pcd = s["pcd.npy"]            # (8192, 3)
    vertices = s["vertices.npy"]  # (V, 3)
    faces = s["faces.npy"]        # (F, 3)
    label = s["json"]["genus"]

With 🤗 datasets:

from datasets import load_dataset

ds = load_dataset("LouisM2001/donut", split="train", streaming=True)

The reduced and crossval splits are id lists in splits/: filter samples on __key__ (or json["id"]).

Metadata

Each sample's .json (and metadata.csv) contains:

• id: unique identifier of the sample
• genus: total number of holes across all mesh components in the sample
• components: total number of connected mesh components in the sample
• sample_code: array of 6 integers describing how many components of each genus are present

Meaning of sample_code

sample_code is an array of 6 integers:

[n0, n1, n2, n3, n4, n5]

Here, ni is the number of mesh components in the sample whose genus is i.

So:

• n0 is the number of genus-0 components
• n1 is the number of genus-1 components
• n2 is the number of genus-2 components
• n3 is the number of genus-3 components
• n4 is the number of genus-4 components
• n5 is the number of genus-5 components

From sample_code, the metadata values are computed as:

genus = sum(i * ni for i in [0, 1, 2, 3, 4, 5])
components = sum(ni for i in [0, 1, 2, 3, 4, 5])

In other words:

• genus is the total number of holes in the full sample
• components is the total number of connected components in the full sample

The distribution of labels in the dataset is shown below.

Examples

sample_code = [2, 1, 0, 0, 0, 0]

This means:

• 2 components of genus 0
• 1 component of genus 1
• total genus = 0 * 2 + 1 * 1 = 1
• total components = 2 + 1 = 3

Summary

DONUT is a dataset of 29,517 samples of manifold 3D structures.

Each sample provides:

• one watertight mesh (vertices, faces)
• one 8192-point cloud
• its topological labels (genus, components, sample_code)

The labels describe the global topology of each sample through its total genus, number of connected components, and component-wise genus distribution.

A previous version of this dataset stored each sample as separate .npz / .npy files in float64. It remains available under the git tag v1-legacy.

Citation

If you use DONUT, please cite the paper FILTR: Extracting Topological Features from Pretrained 3D Models:

@inproceedings{Martinez2026FILTR,
  title={FILTR: Extracting Topological Features from Pretrained 3D Models},
  author={Louis Martinez and Maks Ovsjanikov},
  booktitle={Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR)},
  year={2026},
  url={https://arxiv.org/abs/2604.22334}
}