PDF version of this entire document

Computational Cost

Computational loads are an important factor that has also become a barrier, particularly in 3-D. It is an extension that was described in the previous chapter. The model and NRR assessment methods I have already extended to operate on three-dimensional data. Rather than handle 2-D images, the methods then deal with volumes and - in accordance - shuffle distance neighbourhoods become a box or sphere of voxels, rather than a square or a disk.

There are particular steps in the algorithm whose computational cost is far greater than the remainder. Firstly, one must consider the long time that is required to synthesise many images from appearance models and subsequently use them in an evaluation. The greater the number of synthetic images, the more accurate the results. This relationship means that there is no clear point of balance.

Secondly, the more time-consuming process involves the computation of inter-image distances. With the added complexity of a third dimension, as well as a shuffle distance with large neighbourhood sizes, there is a considerable cost which is proportional to the number of voxels at hand. If image size is increased, it is increased in many dimensions, which slows down the very core of the computation - measuring image distances. At present, speed is relatively satisfactory, but there is much left to be desired, especially due to the scalability of the problem. The previous chapter provided insight into the duration of such experiments, which scale linearly wrt the number of image volumes, of which I have 51.