ENC = VL_FISHER(X, MEANS, COVARIANCES, PRIORS) computes the Fisher vector encoding of the vectors X relative to the Gaussian mixture model with means MEANS, covariances COVARIANCES, and prior mode probabilities PRIORS.
X has one column per data vector (e.g. a SIFT descriptor), and MEANS and COVARIANCES one column per GMM component (covariance matrices are assumed diagonal, hence these are simply the variance of each data dimension). PRIORS has size equal to the number of GMM components. All data must be of the same class, either SINGLE or DOUBLE.
ENC is a vector of the same class of X of size equal to the product of the data dimension and the number of components.
By default, the standard Fisher vector is computed. VL_FISHER() accepts the following options:
If specified, L2 normalize the Fisher vector.
If specified, the signed square root function is applied to ENC before normalization.
If specified, compute the improved variant of the Fisher Vector. This is equivalent to specifying the Normalized and SquareRoot options.
If specified, uses slightly less accurate computations but significantly increase the speed in some cases (particularly with a large number of Gaussian modes).
Increase the verbosity level (may be specified multiple times).