Semantic Segmentation

Introduction

A pixel wise segmentation of drivable path using Fully Convolutional Neural-Network (FCN).

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.

Run

Run the following command to run the project:

python main.py

Discussion

The network implemented in this project is based on this paper.

Training Details

System : AWS EC2 || Instance type : g3.4xlarge

Ubuntu 16.04
Python 3.6+
Tensorflow 1.2+
GPU --> Nvidia Tesla M60 (8GB)
Training time : 40 mins (with following hyperparameters)

---

Model params :

Learning rate : 0.001
Dropout (training keep prob) : 0.5
Epochs : 50
Batch size : 12
Kernel Regularizer : l2_regularizer(1e-3)

Result

Some of the images obtained after training:

False positive sample:

False negative sample:

Future improvements

• Augment training images for better performance
• Use more datapoints (from other similar datasets; cityscape like)
• Increase classes (add pedestrians, vehicles, ...)

Rubric achieved:

1. All the unit test have passed.
2. main.py runs without errors
3. Newest inference images from runs folder included (all images from the most recent run)