Semantic Segmentation using FCN8 model

Introduction

This project labels the pixels of a road in images using a Fully Convolutional 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.

Start

Run

Run the following command to run the project:

python main.py

Experiments

All the Kernals for the end of the encoder, including all decoder layers where init and regularized with the following values.

k_initializer   = tf.truncated_normal_initializer(stddev=1e-2)
k_regularizier  = tf.contrib.layers.l2_regularizer(1e-3)

Run 1

LEARNING_RATE = 1e-4
DROPOUT = 0.5
epochs = 6
batch_size = 5

LOSS: unrecorded 

Run 2

LEARNING_RATE = 1e-4
DROPOUT = 0.5
epochs = 10
batch_size = 1

EPOCH:     9 | BATCH:   288 | LOSS:   0.069465

Final Hyperparameters

LEARNING_RATE = 1e-4
DROPOUT = 0.5
epochs = 10
batch_size = 1

EPOCH:     9 | BATCH:   288 | LOSS:   0.069465

Hardware Used

• AWS Spot Instance <g3.4xlarge> (NVIDIA Tesla M60 GPU)

Reflection

The second run produced sharper results. Looks like increasing the epochs and reducing the batch size did the trick.

Future Todos

• Tackle more labels
• Implement a different encoder model, maybe SqueezeNet?!
• Look into SegNet
• Train on the Cityscapes dataset
• Instance Segmentation