Semantic Segmentation

Introduction

In this project I have developed a Fully Convolutional Network (FCN) which semantically segmented images from the Kitti Road dataset into street and non-street segments.

Project Info

To see the FCN implementation you can have a look at main.py in my CarND-Semantic-Segmentation repository.

Overview

The goal of the project was to implement a Fully Convolutional Network (FCN) using the VGG-16 image classifier architecture to semantically segment(street vs. non-street in images).

Architecture

I used the pre-trained VGG-16 network provided by Udacity and converted it to a Fully Convolutional Network. I replaced the last fully connected layer with a one-by-one convolution and set the depth to the number of classes required.

I have used two skip connections, one-by-one convolutions on layers before and upsampled as a last step.

Optimization Paramters

Loss function: Cross-entropy Optimizer: Adam optimizer

Training

These are the hyperparameters I used:

Input	MSE
keep_prob:	0.5
learning_rate	1e-4
epochs	30
batch_size	8

Results

Loss consistently dropped over the course of the 30 epochs, going from an average of 0.30 in epoch 1 to an average of 0.01 in epoch 30. You can see some of the pictures that have been semantically segmented below:

Conclusion

The FCN was able to achieve consistently good results and most of the time correctly segmented streets against non-streets.

Setup

Frameworks and Packages

Make sure you have the following is installed:

• Python 3
• TensorFlow
• NumPy
• SciPy

Dataset

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.

Start

Implement

Implement the code in the main.py module indicated by the "TODO" comments. The comments indicated with "OPTIONAL" tag are not required to complete.

Run

Run the following command to run the project:

python main.py

Note If running this in Jupyter Notebook system messages, such as those regarding test status, may appear in the terminal rather than the notebook.

Submission

1. Ensure you've passed all the unit tests.
2. Ensure you pass all points on the rubric.
3. Submit the following in a zip file.

• helper.py
• main.py
• project_tests.py
• Newest inference images from runs folder