/Faster_R-CNN

Bootstrapped kentaroy47's code to build a Faster R-CNN in Keras for Object Detection

Primary LanguagePython

Faster R-CNN

This is a Keras implementation of Faster R-CNN. It takes large code snippets from kentaroy47's implementation of Faster R-CNN, which can be found here .

Getting Started

  • Clone the repo locally: git clone git@github.com:jamiejamiebobamie/Faster_R-CNN.git
  • In your terminal, navigate to the main folder of the cloned repo.
  • Install the requirements: pip install -r requirements.txt.
  • You'll need the Kitti Dataset. Download it here .
  • Place the downloaded 'kitti-object-detection' directory in the main project folder.
  • Make a subdirectory: mkdir model_trained
  • Training takes a long time, so download the pickled model from my google drive and place the downloaded model in the 'model_trained' folder.
  • If you wish to train the model yourself, simply ignore the above step.
  • The built model is trained to recognize "Cars" and "Pedestrians".

Your files and folder structure should look like this:

main project folder
├── kentaroy47                      # Python code from kentaroy47's repo.
├── kitti-object-detection          # Downloaded dataset from Kaggle.
│   └── kitti_single               
│       ├── testing
│       │   └── image_2
│       └── training
│           ├── image_2
│           └── label_2
├── model_trained                   # Trained model folder.
│   └── model_frcnn.vgg.hdf5        # Pickled/built model from my Google drive
├── results_images                  # Results images.
├── utils
└── ...[files]...

Prediction

  • To make a prediction type: python3 main.py args in your terminal and press 'enter'. 'args' should be the absolute filepaths of images you wish to make predictions on (separated by spaces).
  • Prediction results are saved in 'results_images'.

Prerequisites

The required packages are listed in the requirements file and are downloaded using the pip install -r requirements.txt command in your terminal.

Authors

  • Jamie McCrory

License

This project is licensed under the MIT License - see the LICENSE.md file for details

Acknowledgments

alt text