PDF version of this document

Current work

Some research and experimentation have been carried out with the intention of achieving good appearance models using warps and group-wise⁸ optimisation techniques. Algorithms have been utilised which are able to manipulate one item of visual data to match another using diffeomorphic warps (round in 2-D or spherical in 3-D). Such process must encompass a large set of data in order to reliably generate a good active appearance model. Cross-validation is used to repetitively evaluate whether the model is altered appreciably or not. If a given warp appears to have the opposite effect, namely increase the difference between the model and the target or even leave it unchanged, it should be then discarded and the search for effective combination of warps virtually backtracks.

Some aspects of this current research are heavily based on the work of Davies et al., but intensity attributes of the data are quite badly handled, especially given the high-performance of other components in the whole active appearance model. No existent evidence indicates that the current solution is the best solution or even a good one; in fact, quite the contrary holds. Better knowledge of the variation of intensity and its dependency on shape needs to be acquired first. It is still unknown whether any real correlation as such exists and, if so, which approach can capture it faithfully. Current approaches base this correlation on experimental evidence. In other words, textures are extracted from the training examples and recorded as a vector which is also statistically reliant on shape.

Another issue that is to some extent open for discussion is the procedural approach of geometrically transforming an image or image space. The methodology, precedence and ordering in which warps should be applied are not obvious. There is some satisfactory evidence though that current work surpasses its predecessors. Diffeomorphic warps and the issues related to them remain beyond the remit of the upcoming research work, yet it is crucial that their properties are fully realised. Since they affect both shape and texture, they have an impact on later stages of AAM optimisation.

Figure 5 can finally be presented as many of its constituent parts have been elaborated on. The images at the top of the figure are not all required to fulfil some predefined conditions (for example, having a large rectangular region at the centre), but some similarity between them is essential if valuable outcomes are sought. The warps applied to these images make group-wise registration possible. Comparison of each image to the reference image is still a valid choice, but empirical evidence suggests that results will then be inferior.

Figure 5: Active Appearance Models optimisation