Weapon Violence Dataset 2.0: A synthetic dataset for violence detection

Area of Science:

• Computer Vision
• Artificial Intelligence
• Machine Learning

Background:

• Acquiring authentic data for sensitive research areas like violence detection is challenging due to ethical concerns and limited access.
• Existing violence detection datasets often rely on non-authentic sources like movies or general online videos, highlighting a critical data scarcity.
• The need for diverse and ethically sourced data is paramount for training robust AI models in critical domains.

Purpose of the Study:

• To introduce the Weapon Violence Dataset (WVD), the first synthetic virtual dataset specifically designed for violence detection.
• To address the limitations of current datasets by providing a novel, ethically sourced, and scalable data resource.
• To facilitate the training of deep learning models for violence detection using synthetic data.

Main Methods:

• Generated synthetic violence scenarios within the photo-realistic video game Grand Theft Auto-V (GTA-V).
• Captured video clips featuring person-to-person fights with various weapons (hot and cold) from a frontal view.
• Created three distinct categories: Hot violence, Cold violence, and No violence (control class).
• Included both normal RGB and optic flow videos for comprehensive model training.

Main Results:

• The Weapon Violence Dataset (WVD) was successfully created and is publicly available on Kaggle.
• The dataset provides a controlled and ethically sound alternative to real-world sensitive data.
• The synthetic nature allows for scalability and augmentation to meet future research demands.

Conclusions:

• The Weapon Violence Dataset (WVD) represents a significant advancement in addressing data scarcity for violence detection research.
• Synthetic data generation using video games offers a viable and ethical approach for sensitive AI applications.
• The WVD is poised to enable the research community to develop more effective violence detection models.