Empirical Analysis on the Effectiveness of Pre-trained Models in the Identification of Physical Violence Against Women in Videos for a Multi-class Approach

Violence against women captured in videos and surveillance systems necessitates effective identification to enable appropriate reactions for controlling and mitigating of its effects in public spaces and the potential apprehension of aggressors. While several algorithms have been developed for violence detection, their evaluation has primarily focused on controlled scenarios with clear differentiation between violent and non-violent scenes, representing two-class identification problems. However, real-world situations often present challenges where specific actions, such as hugs or effusive greetings, fall into an ambiguous class that is difficult to classify. Consequently, this transforms into a multi-class identification problem. In this study, we assess the performance of three pre-trained models, namely VGG16, ResNet50, and InceptionV3, to evaluate their efficacy in addressing the multi-class identification challenges. Furthermore, we compare their performance against datasets consisting of two-class classifications, where the models generally exhibit satisfactory results. Our analysis reveals that the models struggle to differentiate the ambiguous scenes effectively, with Inception V3 achieving a 0% correct detection rate for this class. Notably, VGG16 outperforms the other models, attaining accurate detections of 48% for the ambiguous class, 75% for non-violence scenes, and 54% for violence scenes. This research sheds light on the limitations of current classification models when confronted with the complexity of real-world scenarios. The findings emphasize the importance of developing improved algorithms capable of accurately distinguishing ambiguous situations and enhancing the performance of violence detection systems.