monkey-trainer

Cascade Classifier trainer to generate xml classifiers for #monkey-vision

Training Cascade Classifier

Step 1

Collect the test data, screenshots. Run the classifier.py scripts for a selected mobile platform iOS or Android and use the OpenCV modal preview to trigger screenshot capture at a specific time from the running device. (The emulator/simulator must be running)

• Press "p" for positive screenshot capture
• Press "n" for negative screenshot capture
• Press "q" to quit/cancel the capture

Step 2 [Optional]

Resize collected screenshots. Run the file monkey_compact.py. It will resize all the captured images into width 100px and scaled height. After running this script two new folders will be created /pos and /neg containing the resized images. In the following steps is suggested to use the resized images for faster training.

Step 3

Generate negative description file. The .txt file must contain all the paths of the negative images. To generate this file run the method generate_negative_description_file from '/cascade.py'.

Step 4 [Optional]

Install opencv_annotation, opencv_createsamples & opencv_traincascade, if not installed. https://docs.opencv.org/master/d0/db2/tutorial_macos_install.html

# Steps

git clone https://github.com/opencv/opencv.git

# opencv_createsamples & opencv_traincascade are not part of the latest release/version
# cd into the opencv folder and checkout version 3.4 branch
git checkout 3.4

# cd back to parent dir and create the build directory
mkdir build_opencv
cd build_opencv

cmake -DCMAKE_BUILD_TYPE=Release ../opencv

make -j7 # runs 7 jobs in parallel

Step 5

Generate positive description file by using opencv_annotation. Run command:

/path/to/opencv_annotation --annotations=positive.txt --images=positiveDir/

# Example:

/Users/${user}/Desktop/projects/monkey-cli/build_opencv/bin/opencv_annotation --annotations=positive.txt --images=positive/

Step 6

Create vector file from positives description file. Run command:

/path/to/opencv_createsamples -info positive.txt -w 100 -h 24 -num 1000 -vec positive.vec

# Example
/Users/${user}/Desktop/projects/monkey-cli/build_opencv/bin/opencv_createsamples -info positive.txt -w 100 -h 24 -num 1000 -vec positive.vec

Step 7

Train cascade. https://docs.opencv.org/master/dc/d88/tutorial_traincascade.html

/path/to/opencv_traincascade -data output/  -vec positive.vec -bg negative.txt -w 100 -h 24 -numPos 15 -numNeg 100 -numStages 10


# Example:
/Users/${user}/Desktop/projects/monkey-cli/build_opencv/bin/opencv_traincascade -data output/  -vec positive.vec -bg negative.txt -w 100 -h 24 -numPos 15 -numNeg 100 -numStages 10

# - numPos: must be less than the number of the drawn rectangles created in Step 5.