The approach used could be classified as low-level, because relatively few assumptions are made. Large amounts of data are required in order to proceed. The resulting reconstruction has more detail than a higher level approach, and does not provide as much higher level information. The strategy used sets the 3D reconstruction problem in a statistical framework. The model is deformed so that it moves towards the most likely surface. Collections of range measurements form 'ridges' of probability, with the most likely location of the surface at a given point being associated with the high point of the ridge. The level sets of volumes are deformed to fit a surface to these probability ridges, while maintaining other desirable properties such as smoothness and continuity. The resulting surface created is that which is most likely to have given rise to the data. Level sets are well suited to this approach because of their flexibility in terms of topology, many degrees of freedom, multi-scale representation, and efficient means of calculating geometric surface properties.
| Range surface 1 |
| Range surface 2 | Stage 1 | Stage 2 |
| Range surface 3 |
Download an IV model of the final
result (1.5 MB). (Mesh reduction has been applied to keep the size of
the model down, thereby resulting in a loss of some detail)
There is also a WRL (0.9 MB)version available.