RADNET

Rock Art Detection Network.

Built for usage with Keras and Tensorflow backend.

Code repository that has been based on:

• https://github.com/kbardool/keras-frcnn
• https://github.com/RockyXu66/Faster_RCNN_for_Open_Images_Dataset_Keras

Data

Create a data folder through using the preprocess tool in the sibling folder.

Files

• train.py - Training a model from scratch.
• cont_train.py - Continue to train from an existing model.
• test.py - Evaluate model after training.
• predict.py - Make predictions from a trained model on a single panel.
• test_data.py - Script for exploring training data and see that data processing is working properly.
• faster_rcnn/
  • augmentation.py - Methods for performing data augmentation.
  • config.py - Defines model configuration.
  • losses.py - Includes all losses used when training.
  • RADNet.py - Class for creating the network for prediction.
  • RoiPoolingConv.py - Class for performing RoI Pooling.
  • rpn.py - Methods for creating the RPN network together with helper functions.
  • utils.py - Various utility functions for reading and feeding data.
  • base_models/ - Different base networks for processing images.

Setup

1. Create model folder in root directory.
2. Create environment based on environment.yml.
3. Download the model weights here
4. Place the weights in faster_rcnn/base_models/

Run

1. Setup faster_rcnn/config.py.
2. Run test_data.py to see that data reading/feeding works and get some plots and outputs that could be used for changing the config.
3. Edit test.py and cont_train.py to setup epoch length, nr of epochs and if validation is going to be applied.
4. For running training and testing in background: nohup bash -c 'python train.py; sleep 60; python cont_train.py; sleep 60; python test.py' &