Semantic Segmentation

Dataset Generation and Training Scripts for CAV 2021.

Installation and Required Packages

Matlab 2020a Toolboxes utilized:

Deep Learning Toolbox
Computer Vision Toolbox
Parallel Computing Toolbox

Python Packages:

keras
Numpy
OpenCV
pathlib

If you don't have the above python packages. They can be installed using pip.

$ pip install numpy
$ pip install opencv-python
$ pip install pathlib
$ pip install keras

Install unzip.

sudo apt-get install unzip

Datasets

The dataset that used for our experiments is the M2NIST Dataset. The dataset was created in order to teach the basics of semantic segmentation with convolutional neural networks without requiring the use of complex architechtures that take long to train. More details can be found at the above link. This dataset contains 5000 images with a shape of (64,84,1). Each segmentation mask contains 11 classes (Digits 0-9 and a background label), shape:(64,84,11). Segmented Examples are displayed below.

For most segmentation tasks, 5000 images is not sufficient to obtain reasonable performance. Thus we created an additional 50000 images used to train the network. The original M2NIST dataset is only used for testing. If you wish to generate more images, the script we used to generate images can be found here.

Lower Dimensional M2NIST

We also created a lower dimensional variant of the M2Nist. The images in this dataset have a shape of (28,56,1) and may contain either one or two digits. We generated 50,000 training images and 10,000 testing images. The script used to generate this dataset can be found here.

MNIST MASKS

Finally, we also created a segmentation dataset using MNIST. Each image in this set has a shape of (28,28,1). Segmented Examples are shown below. The script used to generate this dataset can be found here.

Setup

The experiments located in this respository were conducted using a machine with the following specifications: ``` OS: Ubuntu 16.04.6 LTS (Xenial Xerus), 4.15.0-45-generic x86_64 GPU(s): 1 GeForce GTX 1080 ``` To reproduce the results:

Generate the Datasets

To generate all three datasets run the following:

chmod u+x setup.sh && ./setup.sh

Training Models and Evaluating the Models

The training scripts used to train the segmentation models are located in the matlab directory. Execute each of them to train segmentation models on the datasets generated above. We also provide a set of pre-trained networks for each of the aforementioned datasets.

To train the networks in matlab use the following script

CAV2021 Networks

The networks shown in Table 1 were trained using the following scripts. The model files will be stored in the models directory. Each file is labled with the corresponding weighted iou.

N1: matlab/mnist.m | pre-trained model and weights: net_mnist_3_relu.mat
N2: matlab/mnist_mp.m | pre-trained model and weights: net_mnist_3_relu_maxpool.mat
N3: matlab/mnist_dilated.m | pre-trained model and weights: net_mnist_dilated_83iou.mat
N4: matlab/m2nist_dilated_ap.m | pre-trained model and weights: m2nist_62iou_dilatedcnn_avgpool.mat
N5: matlab/m2nist_tranposed_training.m | pre-trained model and weights: net75iou_avgpool.mat
N6: matlab/m2nist_dilated.m | pre-trained model and weights: m2nist_dilated_72iou.mat

Docker

The networks can also be trained through the use of docker. To build the docker container run the following:

$ ./build_docker

This will take 10-15 minutes depending on your machine. Once that completes launch a container by running the following:

$ xhost +local:docker
$ ./run_docker.sh

In the container terminal run the below command. It will prompt you to provide a mathworks account login. Executing the command will begin training the networks. Accessing the model files can be done by editing the docker run command in run_docker.sh to include volume tags.

matlab -r "run('SemanticSegmentation/matlab/train_networks.m')"

verivital/SemanticSegmentation