Estimation of image rotation angle with convolutional neural networks

• link to github repository: https://github.com/tobicar/RotationNet
• created with pycharm IDE
• parts of the code better executable with extension pycharm cell mode

• https://drive.google.com/drive/folders/15beJl1wzXnVS1MxS23P35KZ2HMl9YX1D?usp=sharing
• https://drive.google.com/drive/folders/1krabuviNRljeFoleDO_yUBg43UvG8Xce?usp=sharing

• train_coco_dataset.py
  • This script can be used to train the model on the COCO dataset. The trained model is saved in the models folder and the history in the models_history folder.
• train_street_view_dataset.py
  • This script can be used to train the model on the Street View dataset. The trained model is saved in the models folder and the history in the models_history folder.
• predict_model.py
  • Here a trained model can be loaded and a prediction of an image can be performed. The image can be loaded via a FilePath and the rotation angle for the previously performed rotation can be specified.
• evaluate_dataset.py
  • In this script, a trained model can be loaded and an evaluation of the validation dataset can be performed with the model.
• create_history_plot_from_file.py
  • This script reads the saved history of a trained model and generates a loss and an angle error plot.
• helpers.py
  • This script contains helper functions necessary for training, evaluation and testing. These are on the one hand image rotation functions, custom loss functions and plot functions.
• test_model.py
  • Script to test different rotations of an image with a trained model.
• Files that are no longer used:
  • rotate_and_save_to_file.py
    • File loads images from a specified folder and stores them in a new folder randomly rotated. The target structure is such that one folder is created per rotation angle.
  • generate_rotated_file_structure.py
    • File loads images from a folder where the images are already rotated without structure. The images are moved to a destination folder and a folder is created for each rotation angle.
  • train_rotated_coco_dateset.py
    • File to train a model directly on rotated images.

• data
  • Must be created by downloading the images from Google Drive and then pushing them into the Models folder
• models
  • Contains the two models trained over 100 epochs
• models_history
  • Contains the histories in Numpy format of the trained models