3-D Reconstruction of Multi-object Scenes from Multiple Images

Photorealistic 3-D models are used in a wide variety of applications from entertainment and games, through to simulation, training . Algorithms to automatically create such models from ordinary photographs can vastly reduce the workload and expense associated with acquiring such models. The vast majority of research into reconstructing 3-D models from images has concentrated
on the case of single objects.

This thesis presents a method to model complex multi-object scenes in a series of steps starting with a set of images which surround a scene and ?nally producing a complete photorealistic representation of the objects. The probabilistic space carving algorithm is used to provide an initial estimate of shape as it makes no assumptions about the shape of the scene aside from the bounding cuboid. This representation is smoothed by fitting a Radial Basis Function implicit surface, which smoothes noise and interpolates any missing data. Errors which persist are addressed by a matching surface points between images and estimating the perspective transformation between them which is used to calculate the correct position for the point, which is consistent with the input images. The model may be corrected by constraining the surface to pass through these points. The smoothing properties of RBFs can cause problems by interpolating across objects which are close together, causing them to be joined in the representation. A method is presented to correct this by enforcing consistency between edges in 2-D and 3-D.

Experiments are conducted using real image sequences of complex multi-object scenes. Both qualitative and quantitative evaluations are performed demonstrating the effectiveness of the methods presented. In addition to modelling all of the objects present, colour surfaces are produced from which even the text is legible. A detailed study is undertaken into the factors which influence the effectiveness of techniques to recover partially or fully fused objects and conclusions are drawn which hint at the ultimate limit of accuracy in the case of multiple objects.