Contrails Segmentation

This repository is an attempt to develop a competing solution for this Kaggle competition.

Problem Statment:

This competition on Kaggle is a segmentation challenge. The dataset contains a bunch of satellite images of dimension 256x256. A bunch of such satelite images form a record and each record has a corresponding ground truth image which is a binary mask where 1 denotes presence of a contrail and 0 denotes its absence.

Setup

Please ensure that the following packages has been installed.

• Python3
• Tensorflow 1.12.x
• OpenCV 4.x
• NumPy
• pandas
• tqdm
• imutils

Dataset

If you look at the dataset on Kaggle, you will see that the data is stored in NumPy (.npy) files. These NumPy files are nothing but normalised NumPy arrays. Due to the size of this dataset (approx. 450 GB), I do not recommend downloading the train dataset. Instead, a competing member has already converted those np.float32 arrays into np.uint8. Please click here to download this dataset. Once downloaded, please download the validation dataset as it is from the competition page and extract it. Now we are ready to generate .jpg images from those NumPy arrays. To do so, please run the below script,

python image_generator.py

The above script will convert those .npy arrays into jpg images. This script will store the training data in data directory and validation data in val_data directory. Within these folders, there are two additional folders, viz.

• images - This folder contains the satellite images.
• ground_truth - This folder contains the ground truth for the corresponding satellite images.

Models

In order to successfully compete in this competition, I have developed three different segmentation models. The source code of these models resides inside model package. Following models are developed

1. UNet
2. PSPNet (with ResNet50 as backbone)
3. DeepLabv3 (with UNet encoder as the backbone)

Training

In order to train a model, simply run the following command,

python main.py

Testing

Please download the test dataset from the competition page and extract it. Once extracted, run the below command,

python test.py

Note: Please train a model using the above script before running testing. Once a model has been trained, it will be saved in saved_model directory. The test.py script will load the model from this directory.