Introduction

It is implementation of training module in C++ from article Boosting Sex Identification Performance http://www.cs.cmu.edu/~har/iaai2005.pdf

Datasets

Datasets consist of images with 3 different sizes of man and woman face images.

file: xxsamples00.vec Where xx equal to 16, 20 or 24

And images to the end of the file

All samples are normalized to mean = 128 and variance = 64.

To train model: *.exe --train 1 --data_filename "16samples00.vec" --percent_trainset 0.8 --num_classifiers 20 --save_period 10

To evaluate accuracy of trained model: *.exe --data_filename "16samples00.vec" --model_filename "model.bin" --percent_trainset 0.8

--train <1 or 0> 1 - to train model 0 - to evaluate accuracy of trained model (default)

--data_filename File with images in *.vec format

--model_filename Filename of trained model to evaluate accuracy

--percent_trainset Percent of data as trainset should be in [0.5; 1], to evaluate accuracy [0; 1] (default - 0.0 (0 %)

--percent_classifiers Percent of classifiers to observe in every iterations, should be from 0.01 to 1. (default - 0.8 (80%))

--num_classifiers Amount of classifiers to train. Should be from 1 to 1000.

--save_period Save trained classifiers every 'period' iterations (default - 100)

--balanced <1 or 0> Set flag to 1 if you want balanced train data (default - 0)

file: model_HxW_NC_PTS_B_PC.bin

H, W - height and width of training samples
NC - number of weak classifiers
PTS - persent of dataset for training
B - balanced or unbalanced train set
PC - percent of weak classifiers to choose random from all classifiers on every step

model_HxW_NC_PTS_B_PC.bin

Number of weak classifiers times