We built and trained a Convolutional Neural Network in Tensorflow to segment roads on satellite images,
i.e. assign labels road=1, background=0 to each pixel. We implemented a convolution/deconvolution U-Net with dilated layers from Ronneberger et al. (2015). Data augmentation was a key element for performance improvement, we applied rotations, mirroring extension and vertical/horizontal flips.
The train and evaluation data contains 100 and 50 images of size 400x400 and 604x604 respectively. These are some samples of our predictions on evaluation images:
 |
 |
|---|---|
 |
 |
Consult our report for further information.
This project was part of the Machine Learning course taught at EPFL by Prof. Urbanke and Prof. Jaggi.
Our setup requires a Default Unix Environment (Ubuntu 16.04, or MAC OS X) with an installed Pyhton 3.5 or 3.6 Environment. Our implementation also requires the command sha256sum which is integrated OOB in Ubuntu. If you use OS X or Windows, install the command or do the model verification by computing the SHA256 with the appropriate tools. (Read more about model verification below)
The necessary can easily be setup using pip and the provided requirements.txt.
pip install -r requirements.txtTo generate our final Kaggle submission execute:
./run.pyThe predictions and the submission file are saved in the folder prediction/
To train a new model, you may run:
python3 src/tf_aerial_images.py \
--train_data_dir ./data/training/ \
--save_path=./runs \
--logdir=./logdir \
--num_epoch=25 \
--batch_size=1 \
--patch_size=388 \
--gpu=1 \
--eval_every=1000 \
--stride=12 \
--train_score_every=10000 \
--image_augmentation \
--rotation_angles 15,30,45,60,75 \
--ensemble_prediction \
--dilated_layers \
--num_layers=6 \
--dropout=1.0To inspect the model during and after training, use tensorboard:
tensorboard --logdir=./logdirWhen running tf_aerial_images.py, the following flags may be set to control the application behavior:
| Flag | Description |
|---|---|
| batch_size | Batch size of training instances |
| dilated_layers | Add dilated CNN layers |
| dropout | Probability to keep an input |
| ensemble_prediction | Enable ensemble prediction |
| eval_data_dir | Directory containing eval images |
| eval_every | Number of steps between evaluations |
| eval_train | Evaluate training data |
| gpu | GPU to run the model on |
| image_augmentation | Augment training set of images with transformations |
| interactive | Spawn interactive Tensorflow session |
| logdir | Directory where to write logfiles |
| lr | Initial learning rate |
| model_path | Restore exact model path |
| momentum | Momentum |
| num_epoch | Number of pass on the dataset during training |
| num_eval_images | Number of images to predict for an evaluation |
| num_gpu | Number of available GPUs to run the model on |
| num_layers | Number of layers of the U-Net |
| patch_size | Size of the prediction image |
| pred_batch_size | Batch size of batchwise prediction |
| restore_date | Restore the model from specific date |
| restore_epoch | Restore the model from specific epoch |
| restore_model | Restore the model from previous checkpoint |
| root_size | Number of filters of the first U-Net layer |
| rotation_angles | Rotation angles |
| save_path | Directory where to write checkpoints, overlays and submissions |
| seed | Random seed for reproducibility |
| stride | Sliding delta for patches |
| train_data_dir | Directory containing training images/ groundtruth/ |
| train_score_every | Compute training score after the given number of iterations |