Universal Violence and Fight Detection Benchmark

A 16GB benchmark combines three distinct video datasets: RWF-2000, Real-Life Violence, and Movie scenes. It serves as a unified resource for training and evaluating models on violence and fight detection tasks. The dataset was compiled and hosted on Kaggle.

Use Cases

• Train action recognition models on labeled video clips from the RWF-2000 subset.
• Benchmark fight detection algorithms using the Real-Life Violence scenes.
• Evaluate model generalization across different video sources like movie scenes.
• Develop multi-source training strategies leveraging the combined dataset.

Strengths

• Combines three established datasets into a single 16GB benchmark.
• Includes video data from varied contexts: surveillance, real-life, and movies.

Limitations

• Specific row counts, class balance, and video durations for each subset are unknown.
• The geographic and temporal coverage of the source videos is unspecified.

Provenance

Source

Kaggle

Collection Method

Compilation of existing video datasets (RWF-2000, Real-Life Violence, Movie scenes).