Operator Reference
add_sample_class_gmm (Operator)
add_sample_class_gmm — Add a training sample to the training data of a Gaussian Mixture
Model.
Signature
Description
add_sample_class_gmm adds a training sample to the Gaussian
Mixture Model (GMM) given by GMMHandle. The training
sample is given by Features and ClassID.
Features is the feature vector of the sample, and
consequently must be a real vector of length NumDim, as
specified in create_class_gmm. ClassID is the
class of the sample, an integer between 0 and NumClasses-1
(set in create_class_gmm).
In the special case where the feature vectors are of
integer type, they are lying in the feature space in a grid
with step width 1.0. For example, the RGB feature vectors typically
used for color classification are triples having integer values
between 0 and 255 for each of their components. In fact, there
might be even several feature vectors representing the same point.
When training a GMM with such data, the training algorithm may tend
to align the modeled Gaussians along linearly dependent lines or
planes of data that are parallel to the grid dimensions. If the
number of Centers returned by train_class_gmm is
unusually high, this indicates such a behavior of the algorithm.
The parameter Randomize can be used to handle such
undesired effects. If Randomize > 0.0, random Gaussian
noise with mean 0 and standard deviation Randomize is added
to each component of the training data vectors, and the transformed
training data is stored in the GMM. For values of
Randomize 1.0, the randomized data will look
like small clouds around the grid points, which does not improve the
properties of the data cloud. For values of Randomize
>> 2.0, the randomization might have a too strong
influence on the resulting GMM. For integer feature vectors, a
value of Randomize between 1.5 and 2.0 is recommended,
which transforms the integer data into homogeneous clouds,
without modifying its general form in the feature space. If the
data has been created from integer data by scaling, the same
problem may occur. Here, Randomize must be scaled with the
same scale factor that was used to scale the original data.
Before the GMM can be trained with train_class_gmm,
all training samples must be added to the GMM with
add_sample_class_gmm.
The number of currently stored training samples can be queried with
get_sample_num_class_gmm. Stored training samples can be
read out again with get_sample_class_gmm.
Normally, it is useful to save the training samples in a file with
write_samples_class_gmm to facilitate reusing the samples,
and to facilitate that, if necessary, new training samples can be
added to the data set, and hence to facilitate that a newly
created GMM can be trained anew with the extended data
set.
Execution Information
- Multithreading type: reentrant (runs in parallel with non-exclusive operators).
- Multithreading scope: global (may be called from any thread).
- Processed without parallelization.
This operator modifies the state of the following input parameter:
During execution of this operator, access to the value of this parameter must be synchronized if it is used across multiple threads.
Parameters
GMMHandle (input_control, state is modified) class_gmm → (handle)
GMM handle.
Features (input_control) real-array → (real)
Feature vector of the training sample to be stored.
ClassID (input_control) number → (integer)
Class of the training sample to be stored.
Randomize (input_control) real → (real)
Standard deviation of the Gaussian noise added to the training data.
Default: 0.0
Suggested values: 0.0, 1.5, 2.0
Restriction:
Randomize >= 0.0
Result
If the parameters are valid, the operator
add_sample_class_gmm returns the value 2 (
H_MSG_TRUE)
. If necessary
an exception is raised.
Possible Predecessors
Possible Successors
train_class_gmm,
write_samples_class_gmm
Alternatives
read_samples_class_gmm,
add_samples_image_class_gmm
See also
clear_samples_class_gmm,
get_sample_num_class_gmm,
get_sample_class_gmm
Module
Foundation