Dataset Card for SmallNORB

Dataset Description

The SmallNORB dataset is a real-world stereo image dataset designed for benchmarking algorithms in disentangled representation learning and unsupervised representation learning. It was introduced by LeCun et al. (2004) for evaluating generic object recognition with invariance to pose and lighting.

Unlike synthetic datasets such as dSprites or MPI3D, which are generated as a complete Cartesian product of factors (i.e. every possible combination is present), SmallNORB consists of real photographs of physical toy objects under controlled variations, but not every combination of factors is present — for example, object instances are sampled randomly and the views (azimuth, elevation, lighting) do not form an exact grid.

Each sample contains two views:

Left image (96x96 grayscale)
Right image (96x96 grayscale)

Each image pair is associated with 4 known factors of variation and instance index:

category (object type)
instance (specific object instance)
elevation (camera tilt angle)
azimuth (camera rotation angle)
lighting (lighting condition)

The dataset allows researchers to evaluate representation learning on real-world 3D objects, under complex lighting and pose variations. SmallNORB provides an official train/test split. Typically, instance is not considered as a factor.

Dataset Source

Homepage: https://cs.nyu.edu/~ylclab/data/norb-v1.0-small/
License: other. Small NORB is public domain, for research use.
Paper: Yann LeCun et al. Learning methods for generic object recognition with invariance to pose and lighting. CVPR 2004.

Dataset Structure

Factors	Possible Classes (Indices)	Values
category	0,...,4	airplane=0, car=1, truck=2, human=3, animal=4
instance	0,...,9	specific instance of object
elevation	0,...,8	9 elevation angles
azimuth	0,...,17	azimuth originally 0,2,...,34 → scaled to 0-17
lighting	0,...,5	6 lighting conditions

Note: The dataset is not a complete Cartesian product — instances and views are sampled in the original design. Each sample contains a left image and a right image, both corresponding to the same factors.

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 train set
train_ds = load_dataset("randall-lab/small-norb", split="train", trust_remote_code=True)

# Load test set
# test_ds = load_dataset("randall-lab/small-norb", split="test", trust_remote_code=True)

# Access a sample
example = train_ds[0]
left_image = example["left_image"]
right_image = example["right_image"]
label = example["label"]  # [category, elevation, azimuth, lighting]

# Label breakdown
category = example["category"]      # 0-4
instance = example["instance"]      # 0-9
elevation = example["elevation"]    # 0-8
azimuth = example["azimuth"]        # 0-17
lighting = example["lighting"]      # 0-5

# Visualize
left_image.show()
right_image.show()

print(f"Label (factors): {label}")

If you are using colab, you should update datasets to avoid errors

pip install -U datasets

Citation

@inproceedings{lecun2004learning,
  title={Learning methods for generic object recognition with invariance to pose and lighting},
  author={LeCun, Yann and Huang, Fu Jie and Bottou, Leon},
  booktitle={Proceedings of the 2004 IEEE Computer Society Conference on Computer Vision and Pattern Recognition, 2004. CVPR 2004.},
  volume={2},
  pages={II--104},
  year={2004},
  organization={IEEE}
}