--- title: Udacity Lane Detection Using Semantic Segmentation abstract: | This project implements the "Semantic Segmentation" project required in Semester 3 of the Udacity's "Self Driving Car NanoDegree Program" ---

Context

This project implements the "Semantic Segmentation" project required in Semester 3 of the Udacity's "Self Driving Car NanoDegree Program"

Pre-requisites

The project requires:

1. Pretrained VGG Model: Frozen VGG model, downloadble from here. Look at Readme.md in the folder ’VGGModel’.

2. Kitti Road Dataset: Required for training and testing of the trained model. Downloadable from here. Look at Readme.md in the folder ’KittiDataSet’.

3. Python Dependencies: All python dependencies required for executing the project are listed in the environment.yml. The first step of run.sh installs all required dependencies assuming [ana]conda environment is available. See the section 4 "Running the Project" below.

Note. Automated steps are not fully tested. Watch out for failures.

About This Project

Starting with a pre-trained VGG model, the project add two upscaling (specifically, deconvolution) layers in order to match the size of the network-output to the size of the input image, which is a requirement for semantic segmentation.

The following papers are worth reading about the topic:

• SegNet: A Deep Convolutional Encoder-Decoder Architecture

• Fully Convolutional Networks for Semantic Segmentation

Running the Project

1. Execute run.sh. The script will execute semantic segmentation both for image inputs and video inputs. Comment out steps as required.