license: apache-2.0
Dataset Card for 3dshapes
Dataset Description
The 3dshapes dataset is a synthetic 3D object image dataset designed for benchmarking algorithms in disentangled representation learning and unsupervised representation learning.
It was introduced in the FactorVAE paper [Kim & Mnih, ICML 2018], as one of the standard testbeds for learning interpretable and disentangled latent factors. The dataset consists of images of 3D procedurally generated scenes, where 6 ground-truth independent factors of variation are explicitly controlled:
- Floor color (hue)
- Wall color (hue)
- Object color (hue)
- Object size (scale)
- Object shape (categorical)
- Object orientation (rotation angle)
3dshapes is generated as a full Cartesian product of all factor combinations, making it perfectly suited for systematic evaluation of disentanglement. The dataset contains 480,000 images at a resolution of 64×64 pixels, covering all possible combinations of the 6 factors exactly once. The images are stored in row-major order according to the factor sweep, enabling precise control over factor-based evaluation.
Dataset Source
- Homepage: https://github.com/deepmind/3dshapes-dataset
- License: Apache License 2.0
- Paper: Hyunjik Kim & Andriy Mnih. Disentangling by Factorising. ICML 2018.
Dataset Structure
| Factors | Possible Values |
|---|---|
| floor_color (hue) | 10 values linearly spaced in [0, 1] |
| wall_color (hue) | 10 values linearly spaced in [0, 1] |
| object_color (hue) | 10 values linearly spaced in [0, 1] |
| scale | 8 values linearly spaced in [0.75, 1.25] |
| shape | 4 values: 0, 1, 2, 3 |
| orientation | 15 values linearly spaced in [-30, 30] |
Each image corresponds to a unique combination of these 6 factors. The images are stored in a row-major order (fastest-changing factor is orientation, slowest-changing factor is floor_color).
Why no train/test split?
The 3dshapes dataset does not provide an official train/test split. It is designed for representation learning research, where the goal is to learn disentangled and interpretable latent factors. Since the dataset is a complete Cartesian product of all factor combinations, models typically require access to the full dataset to explore factor-wise variations.
Example Usage
Below is a quick example of how to load this dataset via the Hugging Face Datasets library:
from datasets import load_dataset
# Load the dataset
dataset = load_dataset("randall-lab/shapes3d", split="train", trust_remote_code=True)
# Access a sample from the dataset
example = dataset[0]
image = example["image"]
label = example["label"] # Value labels: [floor_hue, wall_hue, object_hue, scale, shape, orientation]
label_index = example["label_index"] # Index labels: [floor_idx, wall_idx, object_idx, scale_idx, shape_idx, orientation_idx]
# Label Value
floor_value = example["floor"] # 0-1
wall_value = example["wall"] # 0-1
object_value = example["object"] # 0-1
scale_value = example["scale"] # 0.75-1.25
shape_value = example["shape"] # 0,1,2,3
orientation_value = example["orientation"] # -30 - 30
# Label index
floor_idx = example["floor_idx"] # 0-9
wall_idx = example["wall_idx"] # 0-9
object_idx = example["object_idx"] # 0-9
scale_idx = example["scale_idx"] # 0-7
shape_idx = example["shape_idx"] # 0-3
orientation_idx = example["orientation_idx"] # 0-14
image.show() # Display the image
print(f"Label (factor values): {label}")
print(f"Label (factor indices): {label_index}")
If you are using colab, you should update datasets to avoid errors
pip install -U datasets
Citation
@InProceedings{pmlr-v80-kim18b,
title = {Disentangling by Factorising},
author = {Kim, Hyunjik and Mnih, Andriy},
booktitle = {Proceedings of the 35th International Conference on Machine Learning},
pages = {2649--2658},
year = {2018},
editor = {Dy, Jennifer and Krause, Andreas},
volume = {80},
series = {Proceedings of Machine Learning Research},
month = {10--15 Jul},
publisher = {PMLR},
pdf = {http://proceedings.mlr.press/v80/kim18b/kim18b.pdf},
url = {https://proceedings.mlr.press/v80/kim18b.html},
abstract = {We define and address the problem of unsupervised learning of disentangled representations on data generated from independent factors of variation. We propose FactorVAE, a method that disentangles by encouraging the distribution of representations to be factorial and hence independent across the dimensions. We show that it improves upon beta-VAE by providing a better trade-off between disentanglement and reconstruction quality and being more robust to the number of training iterations. Moreover, we highlight the problems of a commonly used disentanglement metric and introduce a new metric that does not suffer from them.}
}