opencv_haartraining(1) train classifier


opencv_haartraining [options]


opencv_haartraining is training the classifier. While it is running, you can already get an impression, whether the classifier will be suitable or if you need to improve the training set and/or parameters.

In the output:

shows the hitrate in the set of training samples (should be equal or near to 1.0 as in stage 0)
indicates the false alarm rate (should reach at least 5*10-6 to be a usable classifier for real world applications)

If one of the above values gets 0 (zero) there is an overflow. In this case the false alarm rate is so low, that further training doesn't make sense anymore, so it can be stopped.


opencv_haartraining supports the following options:

-data dir_name
The directory in which the trained classifier is stored.

-vec vec_file_name
The file name of the positive samples file (e.g. created by the opencv_createsamples(1) utility).

-bg background_file_name
The background description file (the negative sample set). It contains a list of images into which randomly distorted versions of the object are pasted for positive sample generation.

This option is that bgfilename represents a vec file with discrete negatives. The default is not set.

-npos number_of_positive_samples
The number of positive samples used in training of each classifier stage. The default is 2000.

-nneg number_of_negative_samples
The number of negative samples used in training of each classifier stage. The default is 2000.

Reasonable values are -npos 7000 -nneg 3000.

-nstages number_of_stage
The number of stages to be trained. The default is 14.

-nsplits number_of_splits
Determine the weak classifier used in stage classifiers. If the value is
1, then a simple stump classifier is used
>=2, then CART classifier with number_of_splits internal (split) nodes is used
The default is 1.

-mem memory_in_MB
Available memory in MB for precalculation. The more memory you have the faster the training process is. The default is 200.

-sym, -nonsym
Specify whether the object class under training has vertical symmetry or not. Vertical symmetry speeds up training process and reduces memory usage. For instance, frontal faces show off vertical symmetry. The default is -sym.

-minhitrate min_hit_rate
The minimal desired hit rate for each stage classifier. Overall hit rate may be estimated as min_hit_rate^number_of_stages. The default is 0.950000.

-maxfalsealarm max_false_alarm_rate
The maximal desired false alarm rate for each stage classifier. Overall false alarm rate may be estimated as max_false_alarm_rate^number_of_stages. The default is 0.500000.

-weighttrimming weight_trimming
Specifies whether and how much weight trimming should be used. The default is 0.950000. A decent choice is 0.900000.

Specify if initial weights of all samples will be equal.

Select the type of haar features set used in training. BASIC uses only upright features, while CORE uses the full upright feature set and ALL uses the full set of upright and 45 degree rotated feature set. The default is BASIC.
For more information on this see

-h sample_height
The sample height (must have the same value as used during creation). The default is 24.

-w sample_width
The sample width (must have the same value as used during creation). The default is 24.

The type of the applied boosting algorithm. You can choose between Discrete AdaBoost (DAB), Real AdaBoost (RAB), LogitBoost (LB) and Gentle AdaBoost (GAB). The default is GAB.

-err {misclass|gini|entropy}
The type of used error if Discrete AdaBoost (-bt DAB) algorithm is applied. The default is misclass.

-maxtreesplits max_number_of_splits_in_tree_cascade
The maximal number of splits in a tree cascade. The default is 0.

-minpos min_number_of_positive_samples_per_cluster
The minimal number of positive samples per cluster. The default is 500.

The same information is shown, if opencv_haartraining is called without any arguments/options.




This manual page was written by Daniel Leidert <[email protected]> and Nobuhiro Iwamatsu <[email protected]> for the Debian project (but may be used by others).