AI4RISK ViDD - Violence Detection Dataset

The ViDD dataset was produced within the framework of the AI4RISK, contract no. 3Sol(T3)/02.09.2024. The ViDD acronym stands for Violence Detection Dataset. The Ethics Committee of Transilvania University of Brasov, Romania approved the experiments involving human participants. Each participant provided informed consent by signing a GDPR-compliant consent form prior to the recordings. The dataset contains 7,301 video clips, each 2 seconds long, with a resolution of 320 × 240 pixels, a frame rate of 25 fps and stored in MP4 format. The videos were collected using four cameras to obtain four different perspectives of the same area. All faces were blurred and the filenames were randomized using a token consisting of 16 alphanumeric characters to improve privacy. All video clips were annotated independently by two persons and each sequence was carefully verified to ensure correct labeling. The dataset includes four violence categories: shooting (1), throwing (2), punching (3) and running and pushing (4), as well as a non-violence (0) class. The number of video clips per class is: 0 – 4,594, 1 – 624, 2 – 550, 3 – 687, 4 – 648. The main dataset directory contains a CSV file entitle clips_classes listing all clips and their corresponding labels (CSV header: filename, violence_type). The 2-second clips are stored in separate subdirectories by class. Additionally, each class directory contains a CSV file listing the clip names and their associated violence type. This dataset is ideal for a variety of applications and research, including: Violence detection and classification - Training and testing models for automatic identification of violent acts in real-time or post-processing; Smart surveillance systems - Developing solutions for security monitoring in public spaces, schools or other environments; Human behavior analysis - Studying and recognizing complex human actions and interactions; Computer vision research - Exploring neural architectures for action recognition, event detection and anomaly detection in video streams. This work was funded by the Romanian Ministry of Research, Innovation and Digitalization, SOLUTIONS project entitled "Platform for fusion and management of multi-source data collections exploitable by artificial intelligence models for predictive estimation and analysis of risk situations (AI4RISK)”, contract no. 3Sol(T3)/02.09.2024