There have been many developments in the theory of scale space in recent years [28,29,30]. The multi-resolution search using the ASM demonstrates the improvements which can be gained from using of multi-scale techniques. However our current method uses the same shape model (eg the same number of landmarks) at each scale, which appears inefficient. The investigation of the way the shape of objects varies with scale is a rich and interesting field, a proper understanding of which could lead to much more powerful models and search techniques. The intention would be to develop multi-scale flexible shape models whose level of detail increases as the scale at which they are measured decreases. A first step would be to label each example of our training set at each scale, and train a whole pyramid model at once. However, this would be time consuming. We would like to investigate the possibilities of automatically generating suitable landmarks at coarse scales given a set placed at the finest scale. This may involve approaches similar to those used by Morse, Pizer and Liu, [28], Mokhtarian and Mackworth [30] or Griffin et al. [31]. The overall aim would be to develop a method of locating objects more rapidly and robustly. During search, the locations of landmarks found at one scale would give constraints on the positions for landmarks at other scales. We will explore search methods using a coarse to fine approach in which both shape model and grey-level models change as the scale of interest is refined.

To summarise, we will

- examine shape vs scale
- build multi-scale models
- examine search methods