As the subject of computer vision develops, the possibilities of its application to solving problems in other disciplines of computer science becomes apparent. Multimedia is a relatively new and rapidly expanding area of computer science that involves the generation, visualisation/realisation and communication of all types of digital information. Many aspects of multimedia are restricted by a lack of automation or hardware constraints, such as very low bit-rate communications, and it is in these areas that computer vision can provide novel solutions. This thesis presents three vision based techniques that assist the generation and communication of multimedia related material at very low bit-rates. The first is a very high ratio, still image compression method that, given computer vision based image understanding, transfers and generates an image representation from which a user can easily discern the original image content. Semantic and graphics based information allows a user to rapidly view and search a large database of full size, true colour, outdoor images. The second techniques involves the generation of 3-D models from small sets of images. This work avoids expensive equipment and minimises user input in the generation of models for an indoor environment or for much less constrained outdoor scenes, without calibration. The third technique extracts high-level information from statistical models of large sets of images or video sequences. This information can then be used to produce more realistic computer based representations of the real world. The development of each system is described in detail and results from testing and applying the systems are presented.