We have seen great advances in real-time 3D vision in recent years, enabled by algorithmic improvements, the continuing increase in commodity processing power and better camera technology. Research in Monocular SLAM (Simultaneous Localisation and Mapping), where a single agile camera moves through a mostly static scene, was for a long time focused on mapping only enough of a scene to enable robust real-time motion estimation of the camera itself. Attention is now turning however to gradually improving the quality of scene reconstruction which can be achieved in real-time. I will speak about how early work on feature-based SLAM is now being surpassed by methods which aim to map dense scene structure, and how this is leading towards ever-more general 3D scene modelling and understanding.