Click here to view the source images.

In this approach the model is stored as a 3-D volume, with each voxel is assigned a probability value. There are three techniques that can be used to obtain a reprojected image. The most correct approach is to integrate along each ray marginalising the probabilities using Bayes Rule. The next best approach is to use a maximum likelihood (ML) approach. In this case the single best voxel along each ray is selected. The final technique generates a conventional binary voxel array by classifying each voxel as being solid or empty.

It is not easy to apply constraints, such as planarity, in the Space Carving framework. If a rough estimate of the structure is known, for example if the calibration points indicate a small number of planes, small flat voxel arrays can be initialised in the region of these planes. In this example the scene was modeled using two "thick" planes. The inset shows the bay window viewed from a front on position. Notice how the thick texture copes with the change in viewing position.