Research Article

Hand Detection Using Cascade of Softmax Classifiers

Algorithm 1

The procedure of training softmax-based cascade.
Prepare multiclass posture example set and the full-sized background images set . Specify the control factors ,
the stage number , the HOG resolutions for different stages, the posture samples pass-rates for the first stages ,
and the size of train samples . Set the current stage level as , the set of stage-classifiers as . Note that,
all sub-images cropped from full-sized background images are of size in training process.
Train the first stage classifier as follows:
Set , sub-images randomly cropped from images in , and .
Train a softmax model with sample sets and HOG of specified resolution, and modify the model into two SftB
classifiers (Eq. (7)) and (Eq. (8)) based upon the pass-rate .
Add and to . If , go to step . Otherwise, go to step . Here represents the number of
examples in .
Randomly crop sub-image from an image queried from . Add to if . Repeat this process until
reaches to .
Reset and . And go to step .
Train the remaining stage classifiers:
Set example sets and as: , .
Randomly crop from image . Add to if . Repeat this process until reaches to the
predefined .
Train a softmax model with sample sets and HOG of specified resolution.
Then, modify this model into SftB classifiers and based on pass-rate , if .
Add and to , and let .
If , go to . Otherwise, cascade training has been finished and the procedure could be stopped.