---
pretty_name: Moving MNIST (test split)
task_categories:
- video-classification
language:
- en
tags:
- moving-mnist
- video
- synthetic
- grayscale
size_categories:
- 10K<n<100K
---

# Moving MNIST (train split)

This dataset is the classic **Moving MNIST** benchmark set released by Nitish Srivastava et al. for sequence prediction and video representation learning.

- Source file: `mnist_test_seq.npy`
- Original source: https://www.cs.toronto.edu/~nitish/unsupervised_video/
- Content: 10,000 sequences, each 20 frames long
- Frame size: 64x64, grayscale
- Data type: `uint8`

## Dataset structure

This Hugging Face dataset stores one sequence per row:

- Split: `train`
- Number of rows: `10000`
- Feature schema:
  - `video`: `Array3D(shape=(20, 64, 64), dtype='uint8')`

Each `video` item is a full sequence of 20 frames.

## How the original data was created

The original Moving MNIST sequences are synthetic videos formed by placing MNIST digit sprites into a 64x64 canvas and moving them with constant velocity and elastic wall bounces. In this specific benchmark file, each sequence contains **two moving digits** over 20 time steps.

The released benchmark file (`mnist_test_seq.npy`) is arranged as:

- Raw shape: `(20, 10000, 64, 64)` = `(time, sequence, height, width)`

For this Hugging Face conversion, it is reorganized conceptually into per-example rows:

- Per-example shape: `(20, 64, 64)`

## Citation

If you use this dataset, please cite the original paper:

```bibtex
@inproceedings{srivastava2015unsupervised,
  title={Unsupervised Learning of Video Representations using LSTMs},
  author={Srivastava, Nitish and Mansimov, Elman and Salakhutdinov, Ruslan},
  booktitle={International Conference on Machine Learning (ICML)},
  year={2015}
}
```

And optionally reference the project page distributing the benchmark file:

- https://www.cs.toronto.edu/~nitish/unsupervised_video/