README.md

---
license: apache-2.0
task_categories:
- object-detection
language:
- en
pretty_name: Detection Moving MNIST (Easy)
size_categories:
- 100K<n<1M
---


# Detection Moving MNIST (Easy)

|  |  |
|:--------:|:---------:|
| ![annotated_video_0](./annotated_video_0.gif) | ![annotated_video_1](./annotated_video_1.gif) |


### Description

**Repository:** https://github.com/maxploter/detection-moving-mnist


A synthetic video dataset for object detection and tracking, featuring moving MNIST digits with:
- 1-10 digits per sequence
- Linear trajectories with small random translations
- 128x128 resolution grayscale frames
- 20 frames per video sequence
- Digit size 28x28
- Per-frame annotations including:
  - Digit labels (0-9)
  - Center coordinates (x,y)

### Supported Tasks

- Object detection in video
- Multi-object tracking
- Video understanding
- Spatiotemporal modeling

## Structure

### Data Instances

A typical example contains:
```python
{
    'video': [video frames],  # Array of shape (20, 128, 128, 3)
    'targets': [{
        'labels': List[int],          # Digit classes present
        'center_points': List[Tuple], # (x,y) coordinates
    } for each frame]
}
```

### Data Format
- Arrow
- Total dataset size: approximately {PLACEHOLDER} GB
- Frame rate: 10 fps

## Data Splits

| Split  | Size   |
|--------|----------|
| Train  | 60,000  |
| Test   | 10,000   |

## Dataset Creation

### Source Data
- Original MNIST Dataset: http://yann.lecun.com/exdb/mnist/
- Synthetic Generation: Custom Moving MNIST implementation

## Annotations

- Automatically generated during sequence creation
- Includes digit classes and trajectory coordinates

### Simulation Parameters (Easy Mode)
```
{
    "angle": (0, 0),         # No rotation
    "translate": ((-5, 5), (-5, 5)),  # Small random translations
    "scale": (1, 1),         # Fixed size
    "shear": (0, 0),         # No deformation
    "num_digits": (1,2,3,4,5,6,7,8,9,10)  # Variable object count
}
```

## Dataset Statistics

| Statistic                    | Value             |
|------------------------------|-------------------|
| Mean (Train)               | 0.023958550628466375     |
| Standard Deviation (Train) | 0.14140212075592035     |
| Mean (Test)               | 0.024210869560423308     |
| Standard Deviation (Test) | 0.1423791946229605     |

You can check those numbers in the file: [dataset_stats](./dataset_stats.json)

![train_digit_classes](./train_digit_classes.png)
![test_digit_classes](./test_digit_classes.png)

![train_digits_per_frame](./train_digits_per_frame.png)
![test_digits_per_frame](./test_digits_per_frame.png)

## Using the Dataset

### Basic Loading
```python
from datasets import load_dataset
dataset = load_dataset("Max-Ploter/detection-moving-mnist-easy")
```

### Visualization Example
```python
import matplotlib.pyplot as plt
import matplotlib.patches as patches

# Load a single example
example = dataset['train'][0]
frames = example['video']
annotations = example['targets']

# Visualize first frame with bounding boxes
plt.figure(figsize=(8, 8))
plt.imshow(frames[0], cmap='gray')

# Draw bounding boxes
for label, center in zip(annotations[0]['labels'], annotations[0]['center_points']):
    x, y = center
    # Assuming digit size of approximately 28x28 pixels
    rect = patches.Rectangle((x-14, y-14), 28, 28, linewidth=1, 
                             edgecolor='r', facecolor='none')
    plt.gca().add_patch(rect)
    plt.text(x, y-20, str(label), color='white', fontsize=12, 
             bbox=dict(facecolor='red', alpha=0.5))

plt.title('Frame 0 with Object Detection')
plt.axis('off')
plt.show()
```

## Limitations

- Synthetic dataset with simple black backgrounds
- Linear trajectories may not represent complex real-world motion
- No complex occlusion handling or object interactions
- No lighting variations or perspective transformations

## Related Datasets

- Original Moving MNIST: http://www.cs.toronto.edu/~nitish/unsupervised_video/