Re-purposing Machine Learning Models: A Human Action Recognition Model Turned Violence Detector
Letícia Portela, Angel Ayala, Bruno Fernandes, Igor S. Farias, Gervásio Neto, Mouglas Eugênio Nasário Gomes · 2023
In the era of big data, resource utilization is a critical consideration. This is particularly true in machine learning and computer vision, where the complexity of images often necessitates large volumes of data to train models effectively. However, data scarcity presents a significant obstacle in certain areas, such as human action recognition (HAR), particularly in real-world applications like violence detection in videos. This research responds to this challenge by recycling a preexisting human action recognition model to perform a new task - violence detection. The HAR model is used as a feature extractor and combined with classifiers to abstract information from the model and categorize videos into regular activities or violence. This approach allows for adequate performance with less data. Two prominent HAR models, Spatial-Temporal Graph Convolutional Network and Mobile Video Network, pretrained on the Kinetics400 and Kinetics600 datasets, respectively, are repurposed and paired with classifiers for this task. In addition, these models are evaluated on two violence datasets, UCF-Crime and Brazilian Violence, to assess their performance in differentiating between violent and non-violent behavior. Both models achieve high Area Under the Receiver Operating Characteristic scores, indicating their ability to discern between the positive and negative classes.