Fast Indexing of Codebook Vectors Using Dynamic Binary Search Trees with Fat Decision Hyperplanes

Abstract

We describe a new indexing tree system for high dimensional codebook vectors. This indexing system uses a dynamic binary search tree with fat decision hyperplanes. The system is generic, adaptive and can be used as a software com­ ponent in any vector quantization system. The cost of this higher speed (compared to tabular indexing) is a negligible degradation of the distortion error. Neverthe­ less, a parameter allows the user to tradeo# speed for a lower distortion error. A distinctive and attractive feature of this tree indexing system is that it can follow non­stationary codebooks by performing local repairs to its indexing tree. Exper­ imental results show that this indexing system is very fast; it outperforms similar tree indexing systems like TSVQ and K­trees.