In order to stimulate interest and engagement in games, it is important to give players a wide variety of video content that can provide scenario variations each time they engage with the game. However, creating a large volume of diverse video content manually is expensive and time consuming.

This project aims to generate novel video narratives from recorded content with minimal human intervention. This requires automatic visual scene understanding that generates auto tagging of scene content and scene actions, either on a frame-by-frame or short clip basis. As well as understanding frame content, action segmentation strategies will be developed and evaluated. This will enable construction of short novel video narratives - for example, from a manually-defined storyline. Deep learning tools and techniques will be employed throughout this project.