(PRASA, 2010-11) De Villiers, H. A. C.; Niesler, T. R.; Van Zijl, L.
Hand posture reconstruction systems based on large
databases of synthetically rendered images of a 3D hand model
offer a simple and flexible means of exploiting domain knowledge
to provide training data. Such systems may also be applied to
other domains in the posture reconstruction field by changing
the model under consideration.
Typically, the index structures used to answer similarity
queries at runtime explicitly contain the prerendered feature
data. However, the combinatorial explosion resulting from the
multiple degrees of freedom available to the human hand severely
limits the complexity of feature data that may be embedded into
the index structure.
The system presented in this paper exploits real-time objectspace
rendering techniques to rebalance the preprocessing and
runtime workloads such that the space complexity of the database
relative to the number of degrees of freedom is greatly reduced.
A prototype of the database subsystem is implemented and its
properties investigated to obtain insight into its scaling behaviour.