This repo is our participation to the road segmentation project from aicrowd.
The goal of this project is to train a model that can differentiate the road pixels from the background pixels of a satellite image. To achieve this goal, we are provided a training set of a 100 Google Maps images of size 400×400 (see data) and their associated ground truths where a 1 indicates a road and a 0 a background (see data). To evaluate the performance of our model, we are also provided 50 similar unlabelled images of size 608x608 (see data).
We trained a classical U-Net as well as an adapted version of a U-Net. The latter takes as input the 400x400 training images and predicts an image of the same size (using 1 pixel padding during convolutions). To predict the submission images, it directly predicts the 608x608 images.
All necessary pacake are listed in requirements.txt and can be installed using conda with:
conda install --file requirements.txt
To create the final prediction execute:
python run.py
The model is downloaded in run_model, the predicted images are stored in run_pred and the prediction file is called submission at the project root level.
- Nolan Chappuis @Nchappui
- Antoine Masanet @Squalene
- Jonas Blanc @jonasblanc