Performance Analysis of X3D Architecture for Cross-Domain Real-World Violence Detection

Abstract

Real-time violence detection systems need models that are efficient and can adapt to different environments. This study looks at the performance of the X3D-XS architecture and focuses on the issue of generalizing across various domains. The model was trained on three controlled source domains: AVD, HockeyFight, and MovieFight. The model's performance was then tested on the diverse Real-Life Violence Situations (RLVS) dataset. The experimental results show that X3D-XS is highly efficient, achieving inference speeds of up to 191 FPS, which makes it suitable for edge deployment. However, the model faces significant challenges due to domain shift; training on a single domain resulted in varying accuracy between 49.5% and 61.1% on real-world data. This indicates that staged and cinematic violence differ quite a bit from real-life situations. Importantly, combining different source domains improved the model's sensitivity, leading to a Recall of 94.28%. These findings demonstrate that while X3D offers the speed needed for real-time monitoring, relying solely on staged training data is not enough for real-world effectiveness, highlighting the essential need for data diversity in surveillance applications.