Cascade_Toolbox

This is a toolbox to train, test, and run HAAR/LBP cascade easily.

Dependencies

• C++17
• Cmake: sudo apt install cmake
• Zenity: sudo apt install zenity
• OpenCV (versions 2 or 3): sudo apt install libopencv-dev
• Simd (Optional)

How to build?

git clone https://github.com/zanazakaryaie/Cascade_Toolbox.git
cd Cascade_Toolbox
mkdir build && cd build
cmake ..
make

If you want to use Simd library to accelerate the detection, replace the cmake command by:

cmake -DBUILD_SIMD_INFERENCE=ON -DSIMD_LIBRARY_PATH=~/Simd ..

How to use?

Run the generated executables in the build folder following this order:

1. collect_train_data
   Input: a training video (.mp4 or .avi)
   Instructions: draw and edit positive boxes with mouse (no need to draw negatives), press space to move to the next frames
   Output: Positive and Negative folders that contain sample images
   

2. collect_test_data
   Input: a testing video (.mp4 or .avi)
   Instructions: draw and edit positive boxes with mouse, press space to move to the next frames
   Output: a folder that contain .txt annotations of each frame

3. train
   Input: configs.yaml file, Positive and Negative folders
   Instructions: just wait for the training to finish
   Output: trained model in .xml format

4. test
   Input: the testing video, the testing directory (output of 2), and the trained .xml model
   Output: prints the Precision and Recall of the trained detector

5. hard_negative_mine
   Input: the training video, the trained .xml model
   Instructions: right-click in false positives, press space to move to the next frames.
   Output: new negative images added to the Negative folder

6. train
   This is for including hard negatives and training a more accurate model.

7. inference or inference_simd (if linked with Simd library)
   Input: a video (.mp4 or .avi) and the trained .xml model
   Output: the video played with detected objects and FPS
   

Appendix

For more details read my posts on cascade algorithm and training a cascade model