PDF version of this entire document
Spectral masks of the types used so far suffer from a particular weakness that was not foreseen and can possibly be overcome by adjusting some parameters. The short explanation of the weakness is that distances do not increase or energy degraded quite so linearly, meaning that a move along distances in search of a geometrically useful cut will either strip out too much or too little. It makes the mechanism impractical for the recognition purposes at hand. What probably remains needed is a better understanding of which parameters to change and how. We looked at 4 papers related to this, but none of these covered specifically the problem at hand.
After an exchange of code and documentation it was possible to produce some results that look like an improvement.
|
