Browsing by Author "Silva, Paulo Miguel Borges"
Now showing 1 - 1 of 1
Results Per Page
Sort Options
- Analysing the impact of emerging backbones on generalization of video anomaly detection modelsPublication . Silva, Paulo Miguel Borges; Carvalho, Pedro Miguel Machado SoaresVideo anomaly detection plays a crucial role in intelligent surveillance systems, where it is essential to identify events that deviate from normal behaviour. These systems offer several advantages such as real-time monitoring that allows immediate response to security threats, scalability to process large volumes of data across different environments and it ensures anomalies are detected without being influenced by human error or human corruption, highly affected in areas like public spaces and prisons. A significant challenge consists in achieving strong Out-of-Distribution generalization, which ensures models perform effectively on unseen data. This dissertation investigates the impact of emerging backbone architectures and advanced learning techniques on the performance of the models, with a focus on improving their ability to generalize across varied and complex real-world scenarios. The study offers a comprehensive comparison of backbone architectures, ranging from traditional to cutting-edge, for the task of anomaly detection. Additionally, it examines the potential of Self-Supervised Learning methods to overcome the limitations of conventional supervised approaches, particularly in improving generalization across diverse datasets. On the other hand, recent literature on Semi-Supervised models indicates that novel backbones do not show significant improvements. However, leveraging One-Class Classification methods may offer better generalization. The findings reveal that multi-modal self-supervised backbones, such as Contrastive Language-Image Pretraining, demonstrate strong performance in anomaly detection even performing novelty detection, however single-modal techniques like Self-Distillation with No Labels are highly sensitive to scenario conditions. Hybrid architectures like NextViT exhibit limited advancements over existing solutions. Additionally, One-Class Classification methods have proven to be effective in controlled environments with minimal variations, offering a simpler and more robust alternative to complex approaches and backbones.