SmallNORB

Overview

The SmallNORB dataset is a real-world benchmark designed for disentangled and unsupervised representation learning on 3D objects under controlled but realistic variations. It was introduced by LeCun et al. (2004) to study object recognition invariant to pose and lighting.

Each example consists of a stereo image pair, capturing the same object from two horizontally displaced cameras.

Total images (train): 24,300 stereo pairs
Total images (test): 24,300 stereo pairs
Image resolution: 96×96 (grayscale)

Latent Factors of Variation

Each stereo pair is annotated with five discrete factors of variation:

Factor	Discrete Values	Description
`category`	{0, …, 4}	Object category (airplane, car, truck, human, animal)
`instance`	{0, …, 9}	Specific object instance within a category
`elevation`	{0, …, 8}	Camera elevation angle
`azimuth`	{0, …, 17}	Camera azimuth angle (originally 0, 2, …, 34)
`lighting`	{0, …, 5}	Lighting condition index

Each sample corresponds to a single observed configuration of these factors. The dataset does not enumerate all possible combinations, reflecting the constraints of real-world data collection.

Data Structure

When accessing an example using ds[i], you will receive a dictionary with the following keys:

Key	Type	Description
`left_image`	`PIL.Image.Image`	Left view grayscale image (96×96)
`right_image`	`PIL.Image.Image`	Right view grayscale image (96×96)
`label`	`List[int]`	Discrete latent indices: `[category, instance, elevation, azimuth, lighting]`
`category`	`int`	Object category index
`instance`	`int`	Object instance index
`elevation`	`int`	Elevation index
`azimuth`	`int`	Azimuth index
`lighting`	`int`	Lighting condition index

Usage Example

Basic Usage

from stable_datasets.images.smallnorb import SmallNORB

# Load training split
ds = SmallNORB(split="train")
ds_test = SmallNORB(split="test")

# Load both splits
ds_all = SmallNORB(split=None)

sample = ds[0]
print(sample.keys())

left_image = sample["left_image"]
right_image = sample["right_image"]
label = sample["label"]

category, instance, elevation, azimuth, lighting = label

# Optional: make it PyTorch-friendly
ds_torch = ds.with_format("torch")

References

Dataset homepage: https://cs.nyu.edu/~ylclab/data/norb-v1.0-small/
License: Public domain (research use)
Paper: LeCun et al., Learning Methods for Generic Object Recognition with Invariance to Pose and Lighting, CVPR 2004

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},
  volume={2},
  pages={II--104},
  year={2004},
  organization={IEEE}
}