---
language: en
license: apache-2.0 # Or choose another appropriate license
tags:
- video-classification
- violence-detection
---

# Real Life Violence Detection Dataset

This repository contains the "Real Life Violence Dataset" used for training video violence detection models, along with preprocessed versions.

**Associated Model:** [ross-dev/violence-transformer](https://huggingface.co/ross-dev/violence-transformer)
**Original Code Repository:** [ross-sec/violence-transformer](https://github.com/ross-sec/violence-transformer)

## Dataset Description

The dataset consists of video clips categorized into two classes: "Violence" and "NonViolence". It includes real-world scenarios. Due to the nature of the content, user discretion is advised.

## Dataset Structure

This repository contains two main ways to access the data:

1.  **Raw Video Files:**
    *   `Real Life Violence Dataset/Violence/`: Contains `.mp4`, `.avi`, etc., files depicting violent scenarios.
    *   `Real Life Violence Dataset/NonViolence/`: Contains `.mp4`, `.avi`, etc., files depicting non-violent scenarios.

2.  **Preprocessed NumPy Arrays:**
    *   `features.npy`: A NumPy array containing sequences of preprocessed frames. Shape: `(num_videos, sequence_length, height, width, channels)`.
    *   `labels.npy`: A NumPy array containing the corresponding labels (0 for NonViolence, 1 for Violence). Shape: `(num_videos,)`.
    *   `video_files_paths.npy` (Optional): May contain paths to the original video files corresponding to the features/labels.

## Preprocessing (for `.npy` files)

The `features.npy` and `labels.npy` files were generated using the `DatasetCreator` class within the associated model's training script (`train_transformer.py`). The process involves:

*   Extracting `SEQUENCE_LENGTH` (e.g., 16) frames evenly spaced throughout each video.
*   Resizing each frame to `IMAGE_HEIGHT` x `IMAGE_WIDTH` (e.g., 64x64).
*   Normalizing pixel values to the range [0, 1].
*   Skipping videos that couldn't be processed or didn't have enough frames.

Refer to the `DatasetCreator.frames_extraction` method in the [code repository](https://github.com/ross-sec/violence-transformer/blob/main/train_transformer.py) for the exact implementation details.

## How to Use

*   **With Raw Videos:** You can directly load and process the video files from the `Real Life Violence Dataset/` subdirectories using libraries like OpenCV or PyAV for custom preprocessing pipelines.
*   **With Preprocessed Features:** Load the `.npy` files using NumPy:
    ```python
    import numpy as np

    features = np.load("features.npy")
    labels = np.load("labels.npy")

    print("Features shape:", features.shape)
    print("Labels shape:", labels.shape)
    ```
    These arrays can be directly used to create PyTorch or TensorFlow datasets for training models like the associated `VideoTransformer`.

## Citation Information

Please cite the original source of the "Real Life Violence Dataset" if known. If using the preprocessing script or the models from the associated repository, please cite:

```bibtex
@misc{violence_transformer_2025,
  author = {Andre Ross},
  title = {Video Violence Detection Models (MobBiLSTM & Transformer)},
  year = {2025},
  published = {https://github.com/ross-sec/violence-transformer}
}
```

## Disclaimer

The creators of this repository and the associated models are not responsible for the misuse of this dataset. User discretion is strongly advised when working with this data.