DiUS/MaskRCNN-utils

Utilities for working with Mask R-CNN, a neural network for object instance segmentation.

Utilities for working with Mask R-CNN, a neural network for object instance segmentation.

The data preparation pipeline to train a model to segment vein instances using Mask R-CNN is as follows:

Labelbox -> ImageSplitter -> ImageDistributer (test vs rest) -> Augmentor -> ImageDistributer (train vs val)-> MaskRCNN

The steps are explained as follows:

• Use Labelbox to label classes on the image.
• Export the Labelbox project as JSON.
• Generate images their single instance masks using the coordinates in the json:

  python -u labelbox_parser.py \
  --labelbox_json_file path/to/labelbox.json \
  --labelbox_class_names "Sulphide/Partial Sulphide" --labelbox_class_names "Pure Quartz Carbonate" \
  --output_dir path/to/labelbox_parser/output \
  --resize_images
  

• Split the large images and masks from the output above into squares:

  python -u image_splitter.py --input_dir path/to/labebox_parser/output/ --output_dir path/to/image_splitter/output/
  

• Separate 5% of images and masks, splitted above, to be used for testing the trained Mask R-CNN model later and remaining 95% for augmentation for training and validation of new model using:

  python -u separate_test_and_augmentation_images.py \
  --labelbox_output_dir path/to/labelbox_parser/output/ \
  --image_splitter_output_dir path/to/image_splitter/output/with_labels_only \
  --output_test_dir path/to/mrcnn/dataset/stage1_test \
  --output_augmentation_dir path/to/output/augmentation_raw \
  --labelbox_class_names "Sulphide/Partial Sulphide" --labelbox_class_names "Pure Quartz Carbonate"
  

• Apply augmentation to generate more samples:

  python -u augmentation.py \
  --input_dir path/to/augmentation_raw/ \
  --output_dir path/to/augmenation/output/ \
  --number_of_augmented_images_per_original 20 \
  --no-augment_colour
  

• Separate 5% of images and masks, augmented above, to be used as validation dataset and the rest 95% as training dataset:

  python -u separate_train_and_val_images.py \
  --input_dir path/to/augmentation/output \
  --output_dir path/to/mrcnn/dataset/ \
  --labelbox_class_names "Sulphide/Partial Sulphide" --labelbox_class_names "Pure Quartz Carbonate"
  

• Now the path/to/mrcnn/dataset/ contains stage1_train, val and stage1_test dataset directories which can used in Mask R-CNN training and inference.