Semantic Segmentation
In this project, pixels of a road in images are labeled using a Fully Convolutional Network (FCN). I started with the VGG16 network and added skip layers, 1x1 convolutions and upsampling to build the FCN.
This implementation is based on the 2015 paper Fully Convolutional Networks for Semantic Segmentation from UC Berkeley. As for the framework, TensorFlow is used. The network has been trained on an Amazon p2.xlarge GPU instance with around 11GB of graphic memory. The batch size is set to 1 and the network is trained for 10 epochs. With those settings, the results are much better than with a larger batch size and a higher number of epochs.
As regularization, dropout with a drop-rate of 20% is used together with a L2-regularizer for the kernel weights of the decoder convolutions. Adam is used as optimizer, because is a good choice for most deep learning projects. The final loss value is at around 0.05. This project was part of term 3 of the Udacity self-driving car nanodegree program.
Example images:
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. The frozen VGG network is downloaded automatically on the first run and extracted.
Run the following command to execute the project:
python main.py