Extract roads from satellite images

For this problem, a set of satellite/aerial images acquired from Google Maps and the corresponding ground-truth images are provided in /training. Our goal is to train a classifier to segment roads for images in /test_set_image. Our convolutional neural network is provided in run.py and our results using the CNN are in predictions_testing.

This repository contains following parts:

1. Training data — training
2. Testing data — test_set_image
3. CNN code for road segmentation — run.py
4. A pre-computed CNN model — mnist
5. Prediction data — predictions_testing
6. A submission file — tf_submission.csv
7. An image generated in the code for converting the format — sharpen.png

To train your own model, please set RESTORE_MODEL = False. The training might take several hours.

To reproduce the result shown in this file, please set RESTORE_MODEL = True.

This is the work of Fan Zhang, Wenyuan Lv, and Tina Fang.