This is a project I did during my Master program.
This neural network project was created to evaluate the effect of dilated convolutions on a semantic segmentation task. The dataset from the crowdAI Mapping Challenge will be used to test dilated convolutions for binary segmentation. Check out the wiki for further explanations.
The neural network for this project was programmed by using Keras with the Tensorflow backend. Loading and visualizing image data is done with OpenCV. The scripts should be usable if you have installed the above mentioned packages. Keep in mind that this project was designed with the GPU version of Tensorflow and training will be slow or may not work at all with your CPU.
The train.py file can be used to start training if you have the mapping-challenge dataset available and created a configuration file.
A description of the dataset can be found here. The intended training/tests will be conducted with down-sampled images with the shape (150, 150, 3). The dataset is loaded sequentially during training by using the DataLoader. The images/labels are loaded in batches and are randomly flipped or rotated.
Previous training/testing was done on down-sampled versions of the images from the mapping challenge. The images were converted to NumPy arrays with the shape (150, 150, 3) and stored in single .npy files. If you convert the data and labels to numpy arrays you can use the NumpyDataLoader for training.
For loading the training/validation data a .ini file is read. The file should contain the following content:
[DIRECTORIES]
train_data = /path/to/data/train/
val_data = /path/to/data/val/
models = /path/to/models/
logs = /path/to/logs/
predictions = /path/to/predictions/The project is finished and the evaluation can be found here. It is currently not planned to test further architectures in the near future. However I am open for any suggestions for improvements.
Below you can see an example prediction of the UNet architecture. (Left to right: original image, ground truth, prediction)
