Using photographs as the underlying scene primitives provides an inexpensive and realistic environment representation. However, if goals other than pure visualisation are important such as extensive walkthroughs in large scale sites and spatial reasoning or editing then an underlying geometric model that can be directly rendered from arbitrary views is required.
The problem of automatic reconstruction of 3D models has attracted considerable research interest, initially focused on modelling small objects and more recently larger scale environments. The form of these models can be polygonal, volumetric, based on higher order surface patches or CAD models. The geometric representation used depends on both the type of acquisition sensor and the required use of the resulted model. Photorealism is enhanced by direct mapping of the images of the scene as texture onto the corresponding geometric primitives.
Geometry modelling methodologies can be categorised according to the amount of 3D information available and the type of the adopted sensors. Both active and passive sensors have been utilised to provide dense or sparse 3D measurements of the real scenes geometry. Unlike image based techniques, as the level of detail for a scene can be unrestrictedly complex, the computational requirements of the reconstruction process increases because of the bigger and more accurate data sets needed. Despite this principal limitation though, examples of reconstructions on complex real world environment have been achieved and constant advances in both sensor and computer hardware technology are likely to provide the means for arbitrary scene modelling in the future.
An overview of several different geometry based methodologies is presented in this section starting with techniques that utilise dense depth maps obtained from active sensors. Approaches that use the same kind of input data but instead based on camera sensors are subsequently presented. Using again passive sensing two categories of volumetric methods, shape from silhouettes and reconstruction based on the photo consistency constraint are used. Finally research has also addressed modelling from sparse data which is more closely related to this work. Despite the focus being on reconstructing large scale scenes, the feasibility of techniques that extend existing small object modelling approaches is also assessed.