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.

• 16x16 - about 1136 images
• 20x20 - about 4505 images
• 24x24 - about 1139 images

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

• 1 byte - width of dataset (w)
• 1 byte - height of dataset (h)

And images to the end of the file

• w*h bytes - image (first row - last row of image)
• 1 byte - gender (1 for man and 255 for woman)

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

Usage

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)

Model file

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

• 4 byte - Int - number of weak classifiers

Number of weak classifiers times

• 4 byte - Int - first pixel row
• 4 byte - Int - first pixel column
• 4 byte - Int - second pixel row
• 4 byte - Int - second pixel column
• 8 byte - Double - Log(1/betta)
• 4 byte - Int - feature type of weak classifier
• 4 byte - Int - feature parity of weak classifier
• 8 byte - Double - accuracy on train set after adding this weak classifier
• 8 byte - Double - accuracy on test set after adding this weak classifier