In this project, the pixels of the road of some images will be labeled using a Fully Convolutional Network (FCN).
Make sure you have the following is installed:
Download the Kitti Road dataset from here. Extract the dataset in the data
folder. This will create the folder data_road
with all the training a test images.
Run the following command to run the project:
python main.py
The final model uses a batch size of 2, learning rate of 1e-4, keep probability of .5 and runs for 10 epochs. The loss decreases over time, probably running for more epochs will improve the results. It took 38 minutes in a g3.4xlarge instance from Amazon with a NVIDIA Tesla M60 GPU to train.
It uses Adam Optimizer and L2 regularization. On the second epoch results looked like this:
But it got much better (epoch number 10):