Image Segmentation Model for the xView2 Competition

Dual-Image Input U-Net Model Used:

Environment Setup:

• The code in this repository was run in a Python 3.7 environment.
• Project requirements are located in the requirements.txt file

Instructions:

Download the data from:

• https://xview2.org/dataset

Run transform_data.sh

• To consolidate the training and tier 3 data and the testing and holdout data into train and test directories, respectively

Run data_preprocessing.py

• Reads the json label data and creates a ‘targets’ directory containing the pre- and post-disaster image masks

Run train_model.py

• This script imports keras_dataset.py and dual_image_unet.py to use during training.
• LabeledImageDataset takes in the pre and post images along with the post labels - make sure the pattern variable is in the right directory
• U-Net is the dual image input U-Net - make sure the number of output classes is correct (should be 5, for the different levels of damage)
• Best model weights will be saved based on the validation loss. Saved to an hdf5 file.

Run test_process.py

• This script formats the test targets directory from earlier in a way that each target label is of shape (1024, 1024) and each pixel has the class label.
• It also loads in the pretrained model weights and predicts on the test images.
• These predictions are output in the same shape and format as the target labels - into the predictions directory.

Run the final metrics calculator at:

• https://github.com/DIUx-xView/xView2_scoring/blob/master/xview2_metrics.py

Plotting_sample_images.py

• Contains functions needed to plot the masks over the original image. This assists in visualizing the results.