[prev] [index] [next]

Why VA-Files Don't Work (for QBIC)

In the experiments reported by [Weber and Blott] they use uniform data distributions and achieve 0.0005 filtering.

In experiments with QBIC colour histograms, which have extremely skew distributions, we were unable to achieve better than 0.10 filtering on average.

And this was after "tuning" partition boundaries to place approximately equal numbers of points in each partition.

Before tuning, filtering factor was 0.70.

Conclusion VA-file performance is very sensitive to data distribution.

VA-files will also perform poorly if TD is large   (we compute D for every signature)


[prev] [index] [next]