Results from projective geometry have lead to fruitful work in Computer Vision [33,34]. Recently considerable interest has been aroused in what can be deduced from images taken by uncalibrated cameras [35,36,40,41]. A strongly related field is the study of geometric constructions which are invariant to camera parameters and pose [42]. Much of this work has assumed that the objects viewed are rigid, or can only deform in ways which are locally affine [39]. We intend to investigate the generation of flexible models of 3D objects from uncalibrated images, where the objects can deform in more general ways. This would involve building statistical models in a projective space. Early exploratory studies show that this is possible, though care must be taken in choice of reference frame and the statistical models used. Our long term goal is to build models from sets of uncalibrated 2-D monocular images. However, in the shorter term we will work with uncalibrated stereo images, from which some 3D structure can be obtained using Faugeras' methods [35]. We will go on to develop ways of using these models to locate and classify objects in new images. Blake et al [37] and Beardesly et al [38] have developed geometric models which allow affine deformations for tracking objects in image sequences. Related methods may allow the more general flexible models to locate or track objects.

To summarise, the intention is to investigate

- methods of building flexible projective models
- techniques for using such models to locate objects in new images